Artificial Intelligence to Facilitate the Conceptual Stage of Interior Space Design: Conditional Generative Adversarial Network-Supported Long-Term Care Space Floor Plan Design of Retirement Home Buildings

ABSTRACT

This study uses Conditional Generative Adversarial Network (CGAN) to construct a method for generating floor plans for long-term care spaces in retirement home buildings to assist architects in improving interior space design. The results of this study show the following: (1) For the interior design of long-term care spaces in retirement home buildings, the CGAN model has strong understanding and calculation capabilities. The zoning layout of long-term care spaces in retirement home buildings has been completed, and the results show that the CGAN model has reference value. (2) Although there are several differences in the design of CGANs and authentic design, there are still many similarities. Some unreasonable results, such as space generation in corridors and elevator shafts, require further manual correction. (3) According to a later questionnaire survey on the satisfaction of architects and CGAN model design solutions, the difference between the two is not large, which also illustrates the great potential of CGANs for intervention in interior space design. This helps architects create more detailed plans based on the model, greatly increasing work efficiency. Moreover, additional interior space design possibilities can be explored, and to some extent, the architect’s subjective assumptions can also be corrected.

Introduction

Research Background: The Promotion of the Aging Phenomenon and the Importance of Elderly Care Spaces

The trend of the global population aging has become a social phenomenon that cannot be ignored. According to the United Nations, it is expected that by 2050, the global population aged 60 and above will account for 22% of the population (Bloom et al. 2011; Cohen 2003). In this global context, China’s aging problem is particularly prominent. According to the National Bureau of Statistics, China’s elderly population over 60 years old reached 240 million in 2017. This number is expected to increase to 487 million by 2050, accounting for 34.9% of the total population and up from 17.3% (Mao et al. 2020).

There are multiple factors behind this trend. First, China’s one-child policy implemented in the 1980s led to changes in family structure, causing most modern families to have only children (Feng, Poston, and Wang 2014; Rosenberg and Jing 1996). This policy not only exacerbates the aging problem but also leads to the emergence of the “421” family structure, in which one young person is responsible for the care of two elderly people and four grandparents. One study estimated that in 2000, China had 15.6 elderly people out of every 100 working-age people, and this number will increase to 48.5 by 2050 (Su, Hu, and Peng 2017). Older care places a great deal of burden on children, not only in terms of economics but also in terms of the time cost required to care for elderly people, which is the biggest problem faced by children in modern society. Children’s work and life pressures are increasing daily (Hongyan 2003; Ying 2003). How to reduce the burden on family members and medical staff while ensuring the safety of elderly people and how to better meet their psychological needs are top priorities for every designer. To design and optimize interior spaces of retirement home buildings for long-term care are particularly important. Second, with the rapid development of the economy and the improvement of living standards, the lifespan of elderly people has also increased, which has undoubtedly increased the demand for elderly care services. However, the current supply of elderly care services is far from meeting this growing demand. Due to limited land and human resources, the quantity and quality of retirement homes and retirement apartments face considerable challenges, especially in urban areas (Lam and Yan 2022).

As the elderly population continues to grow, many physical and psychological problems among elderly people have gradually become apparent. Physiologically, due to the compromised immune systems of elderly people, most elderly people suffer from one or even multiple chronic diseases (Fulop et al. 2010). Psychologically, elderly people seldom go out for activities due to difficulties in their motor skills and have less and less communication with the outside world, making it difficult for them to accept the rapidly changing life of the outside world. Therefore, their character gradually becomes withdrawn and paranoid, and they are also prone to depression and loneliness and are emotionally sensitive and fragile, thus accelerating aging and death (Lowenthal 1964).

Under these circumstances, local governments began to take action and introduce a series of policies to encourage and support the development of elderly care services. For example, many cities in the Chinese mainland have introduced policies such as financial subsidies, land use concessions, and tax exemptions to promote the construction and operation of retirement homes and retirement apartments (Feng et al. 2020; Liu et al. 2018). Although these policies have alleviated the pressure on elderly care services to a certain extent, they still cannot fundamentally solve the problem. Therefore, how to effectively plan the interior space of retirement homes and retirement apartments to meet the growing demand has become an urgent and important issue. This is not only a social problem but also a design problem. The interior space design of traditional retirement homes and retirement apartments for long-term care often ignores the actual needs of elderly people, such as reduced mobility, reduced vision, memory loss, and emotional instability. These issues need to be fully considered in the design process.

Literature Review

Traditional Method for Interior Design of Long-Term Care Space

The traditional approach to interior design of long-term care space in retirement home buildings is deeply rooted in the experience and intuition of the designer. Far from being simple “subjective experiences,” these experiences are actually tacit knowledge that is difficult to explicitly label on a machine. They represent the designer’s accumulation of years of practice and contain rich insights and creativity. When dealing with the special needs of older adults, such as mobility, physical characteristics, and mental health, designers use this tacit knowledge to solve complex problems. For example, in response to the decline in vision and hearing in elderly people, designers may consider elements such as color contrast, scene accessibility, and spatial cues, which help elderly people better locate and identify the environment (Yu and Wang 2020; Yu, Spenko, and Dubowsky 2003).

The work of traditional designers is not only based on experience but also includes using divergent thinking to create diverse and high-quality design solutions. This strategy frees designers from the constraints of a systematic approach to think and innovate freely. This creativity and flexibility of the human brain can quickly lead to the conception and interpretation of design solutions in 3D space and even more dimensions. The design process is indeed time-consuming, especially when creating a design that is both suitable and will withstand the test of time. This design process generally consists of five stages: the design requirements stage, the conceptual design stage, the design deepening stage, the construction design stage, and the operation and maintenance stage (Chen and Chaudhary 2022; Rahmanifard and Plaksina 2019). Initially, the design requirements stage establishes the functional and capacity needs, laying the groundwork for the project. Following this, the conceptual design stage involves the generation of external and internal plans, which include architectural organization, room layouts, and floor plans, translating the requirements into a tangible vision. Subsequently, the design deepening stage delves into the specifics, focusing on the detailed design of technical equipment, interior finishes, lighting, seating, signage, and other intricate details, refining the project’s overall aesthetic and functionality. The construction design stage then transform these detailed plans into actionable blueprints, leading to the physical construction of the space. Finally, the operation and maintenance stage ensures the designed space not only meets the initial specifications but also remains functional and comfortable for its residents over time.

However, traditional methods also face several challenges, which can be traced back to the earlier stages, particularly in the conceptual design stage. By relying too much on personal experience, certain key factors may be overlooked, resulting in limited design options. This limitation can ripple through an entire project. For example, the stagnation of innovation caused by the long iteration cycles and a lack of diverse design solutions are also common problems that reduce the overall design efficiency (Cui and Chung 2023; Dong, Guo, and Jiang 2014). Despite this, traditional design methods still occupy an irreplaceable position in interior design of long-term care space, especially in considering the specific needs of elderly people and creating a humane and caring environment.

Architectural Design Applications Combined with Algorithmic Technology

With the development of computer technology, the field of architectural space design is undergoing a transformation, especially with the introduction of parametric design tools and machine learning algorithms. The application of these technologies has greatly improved the systematicity and efficiency of the design process. Khean et al. (2018) showed how, by using parametric modeling platforms such as Grasshopper, designers can quickly test and modify various design solutions to achieve efficient iteration of the design process. These platforms support designers in exploring and refining design solutions in a shorter period of time, greatly improving the flexibility and response speed of the design process. Optimization techniques, such as genetic algorithms, particle swarming algorithms, and simulated annealing algorithms, have been explored for automated design solution generation (Jovanović et al. 2020). These technologies automatically generate design solutions through algorithms, significantly reducing the designer’s workload. For example, Kirimtat et al. (2019) showed in their study how these optimization techniques can be used to generate algorithmic models on the Rhino Grasshopper platform to improve architectural design.

However, there are still challenges in applying these advanced technologies to the interior design of long-term care spaces. These methods usually require designers to preset detailed design rules or parameters, requiring designers to have deep professional knowledge and experience (Labib 2022). In addition, there are challenges in the optimization algorithm itself, such as the difficulty in adjusting hyperparameters and the risk of falling into local optima. While solving these problems, machine learning, especially deep learning networks, began to be applied to analyze a large number of design cases to learn the design rules. These methods provide a new perspective on design through a data-driven approach and help solve the limitations of traditional methods in dealing with complex factors (Preisinger and Heimrath 2014). Advanced algorithms, such as deep reinforcement learning, have also been developed to optimize the design decision-making process (Kakooee 2024). Parametric and vector image models have the advantage of ensuring design consistency and compliance with building codes. But this model often relies on strictly defined rules and parameters, which makes it rigid when adapting to rapidly changing design requirements (Zheng and Bingren 2006). When facing changes in requirements or new design trends, parametric models may need to undergo complex modifications or completely reconstruct parameter rules, which is not only time-consuming but also inefficient (Anderl and Mendgen 1995). Furthermore, parametric methods often require in-depth expertise and long-term adjustments when dealing with complex design problems, which limits their application in rapid iteration and innovative design.

In contrast, using CGAN models and methods based on pixel images has obvious advantages: (1) Pixel images are currently the most widely circulated and used data format, which makes them extremely convenient in terms of data collection and model application (Miano 1999). The wide availability of pixel maps means that we can easily obtain a large number of training samples, thereby improving the training efficiency and output quality of the model. (2) The pixel-based CGAN model can be more flexible and faster when processing image data (Mishra and Herrmann 2021). This model can be quickly iterated and updated as user needs change without requiring large-scale modifications to the model architecture (Skorokhodov, Ignatyev, and Elhoseiny 2021). This is particularly important in an environment where nursing home design needs are changing rapidly, as design standards and user needs may be rapidly adapted due to policy changes, technological advances, or changes in market demand. (3) The CGAN model is able to generate highly realistic images by learning patterns in large amounts of data sets (Alqahtani, Kavakli-Thorne, and Kumar 2021; Sampath et al. 2021). This not only helps generate innovative design solutions but also provides intuitive visual feedback in the early stages of design, helping designers and stakeholders better understand and evaluate design solutions. While parametric and rule-based models have advantages in terms of generation accuracy and compliance with specific design specifications, they may not be as innovative and flexible as pixel-based CGAN models. (4) Due to its data-driven learning mechanism, the CGAN model can learn and simulate complex design rules and aesthetic principles from a large number of existing designs (Li et al. 2020). Not only can this model quickly adapt to new design needs, but it can also automatically optimize and update design solutions without frequent manual intervention. This flexibility of CGAN makes it particularly suitable for the design field, especially in architectural design practices that require constant innovation and adaptation to new trends. In addition, the images generated by CGAN are intuitive and detailed, which can provide designers and clients with a clear visual reference, which is extremely valuable during the design communication and program evaluation stages.

At present, machine learning research specifically focused on interior design of long-term care space is relatively limited. Most related research is still in the theoretical stage, and the practical application of these methods need to be verified. For example, the lack of large-scale training datasets suitable for interior design of long-term care space limits the effectiveness of network training (Petersen and Heimrath 2014). Furthermore, evaluating the quality of the generated results is challenging because a balance between functionality and aesthetics is required (Yi, 2017). Although parametric models are still indispensable in some professional applications, CGAN provides a more flexible and efficient design tool that is particularly suitable for responding to the rapidly changing and highly personalized needs in the current field of architectural design. This model’s application potential is particularly prominent in specific fields, such as nursing home architectural design, where it can effectively promote design innovation and efficiency improvements. With the continuous advancement of design practices, these technologies are expected to bring more humane, efficient, and innovative solutions to the interior design of long-term care spaces.

Application of CGAN in Interior Design

According to certain experts, the current automatic floor plan layout-generating processes in architectural design can be categorized into three distinct methods (Weber, Mueller, and Reinhart 2022). A bottom-up strategy entails the process of assembling individual components, such as rooms or preassembled sections, to form a larger structure. As a tool for exploration, it has the ability to rapidly produce several design alternatives. Aggregation strategies can be additionally integrated with heuristics to provide guidance for assembly. Nevertheless, when confronted with extensive design areas, managing intricate constraints or boundary conditions can prove to be exceedingly difficult. A top-down technique can provide an alternative by directly considering geometric restrictions, such as breaking buildings or site boundaries into smaller parts. To do this, several segmentation or packaging tactics can be utilized. Furthermore, ongoing research has focused on utilizing existing buildings and datasets through reference methodologies. The geometric characteristics of current or premade designs can be modified or adjusted to suit new surroundings. Machine learning algorithms are also used as a method to capture graphic or bitmap images that represent the layout of a space’s floor plan. These representations are subsequently encoded into neural networks, allowing for efficient lookups and synthesis (Weber, Mueller, and Reinhart 2022).

In recent years, the application of conditional generative adversarial networks (CGANs) in the field of architectural design has received much attention (Nguyen, Reiter, and Rigo 2014). CGANs, as advanced machine learning tools, can create new possibilities in architectural design. The core advantage of this technology is its ability to learn features from a large number of design cases and generate new design solutions based on these learned features. For example, CGANs have been used to construct architectural facade designs (Abu-Srhan, Abushariah, and Al-Kadi 2022; Lin et al. 2023; Sun, Zhou, and Han 2022; Xueying et al. 2021; Zhang et al. 2022), floor plan designs (Jian and Chen 2020), and interior designs (Park and Kim 2021; Tanasra et al. 2023). Other scholars have proposed a method for generating synthetically labeled floor plan images to train deep learning models. This method generates a floor plan with dimension lines, textual room information, and common objects in context (COCO) object detection labels (Brauksiepe et al. 2023). This approach not only efficiently generates design images but also infers underlying design rules and styles, providing architects with new sources of inspiration. As a machine learning method applied to generative design, the main advantage of CGANs is that they can learn the features of a large number of design cases and generate new design schemes based on them. Compared with traditional computer-aided design systems that rely on manual construction rules, CGANs have stronger learning ability and better generalization performance (Min J, et al. 2023). These studies have focused mainly on the field of residential building design.

In the field of residential architectural design, the application of CGANs is particularly prominent. Research by Grohnfeldt, Schmitt, and Zhu (2018) showed that CGANs can not only generate traditional 2D design drawings but also create 3D models and even complex multidimensional design solutions. This capability greatly expands the boundaries of architectural design, allowing designers to explore more innovative design concepts in a virtual environment.

Interior design for long-term care spaces is a highly specialized field that requires designers to fully consider the physiological and psychological needs of the elderly (DeMello 2016; Fleming and Purandare 2010; Kane 2001). For example, in nursing homes’ floor plan design, accommodation spaces, corridors, support spaces (such as medical rooms and activity rooms), elevators, and public areas must all follow specific circulation design patterns to ensure accessibility and convenience of use. These spatial layouts not only need to meet functional and safety requirements but also consider the mobility and social interaction needs of the elderly. In this context, CGAN’s application offers an innovative perspective. CGAN’s capability lies in its ability to generate outputs under specific conditional constraints, which is critical for creating architectural floor plans that adhere to the interior design of long-term care space standards and user habits. With CGAN, designers can rapidly iterate through diverse design proposals while ensuring compliance and practicality. Moreover, by learning from a large number of design samples, CGAN can reveal hidden patterns and best practices in the interior design of long-term care spaces, thereby promoting a balance between standardization and personalization of designs. The academic community has recognized the potential of applying machine learning in the field of architectural design (Perarnau et al. 2016).

It should be noted that the application of CGAN and other AI tools is to enhance the capabilities of designers, not to replace them. Designers’ professional knowledge and innovative thinking still play an irreplaceable core role in the entire design process. (1) The application of CGAN models and other artificial intelligence tools does not mean reducing designers’ need for professional knowledge (Anantrasirichai and Bull 2022). Instead, effective use of these tools is based on a deep understanding of design principles, building codes, and user needs. When the CGAN model generates design solutions, its input data sets need to be carefully planned and managed, which itself requires designers to have solid professional knowledge and rich experience (Douzas and Bacao 2018; Fan et al. 2020). (2) The role of the CGAN model is to assist designers in exploring innovative design solutions more quickly and effectively, but it does not replace the designer’s creativity and professional judgment. Through AI tools, designers can quickly generate multiple solutions in the early design stage, thereby having more time and resources to deepen the most potential designs and improve the design quality and efficiency of the entire project. (3) The designer’s expertise is crucial for evaluating and selecting design solutions generated by CGAN. Although CGAN can propose design solutions based on learned data patterns, the final screening, optimization, and implementation still require the designer’s professional judgment to ensure the practicability and feasibility of the solution. (4) The CGAN model can also be used as a training and educational tool to help designers better understand design patterns and trends and enhance their design capabilities and innovative thinking. By interacting with advanced technologies, designers can not only improve their professional skills but also better cope with complex and changing problems during the design process.

In summary, the application of CGANs in the field of architectural design provides designers with new tools and perspectives, which can not only improve design efficiency but also provide support for design innovation. As this technology continues to develop and improve, it is expected that CGANs will play an increasingly important role in architectural design, especially in interior design of long-term care space. Additionally, considering the increasing trends of aging populations in China and globally and the growing need for efficient and elderly-specific interior design of long-term care space solutions, this research explores the application of CGAN in the interior design of long-term care space, aiming to advance technological progress and innovation in this field and optimize the design process. Future research should explore how to effectively integrate CGANs into the traditional process of interior design of long-term care space to achieve more comprehensive and balanced design results. Therefore, the research in this article has important academic and practical significance.

Problem Statement and Objectives

In the pursuit of architectural humanization, the interior design of long-term care space presents unique challenges that demand a shift from traditional methodologies toward a more integrated, agile approach that combines evidence-based and rapid prototyping to diversify design solutions and improve overall design efficiency.

The literature review elucidates the constraints inherent in conventional and early computer-aided design approaches, which often fail to effectively address the need for rapid prototyping that allows for diversified design solutions and design efficiency improvement. It is within these limitations that this study situates its exploration, particularly in the emerging domain where machine learning and CGANs have shown notable strengths in architectural design.

Recent advances in architectural design demonstrate the potential of machine learning, especially in the views of the latest research in the field of architecture which emphasized the advantages of machine learning in stimulating new ideas and design iteration through machine learning tools for early-stage design (Horvath and Pouliou 2024). This study draws on this view and hopes to extend the application of machine learning tools to the specific process of retirement home design through CGAN. This study posits that CGAN, as a progressive computer-aided design tool, could significantly contribute to the interior design of long-term care space by enhancing the creativity and flexibility of design proposals in the conceptual design stage. This article explores the use of CGANs to assist architects in the floor plan design of long-term care spaces in retirement home buildings and to improve the design process of architects. The application of CGANs could catalyze a synergistic relationship between designers’ expertise and advanced computational methods, leading to outcomes that are not only innovative but also responsive to the nuanced needs of elderly care environments.

A holistic approach to the interior design process in retirement home buildings serves as the foundation for this research. The paper acknowledges the multifaceted stages of conceptualization, wherein CGAN was introduced as an intervention tool that interacts dynamically with traditional design workflows. As put forward by the previous research, the envisioned role of CGAN is to augment it by mimicking the steps taken by architects during the conceptual design phase and offering a spectrum of design possibilities grounded in the wisdom captured from exemplary architectural works (Rahbar et al. 2019).

In this study, CGAN is trained with a specific dataset to mimic the design steps of human designers in the conceptual design stage. Our contributions are as the following:

·Build a floor plan dataset of long-term care spaces for machine learning.

·Train CGAN for space allocation and objects arrangement.

·Test the two models of CGAN mimicking human designer’s conceptual design process to obtain diversified conceptual solutions.

By doing so, it is hypothesized that CGAN can provide a repository of visualized suggestions based on learned patterns from existing outstanding buildings, which an architect can then refine and align with quality standards and user-centric objectives. The resulting designs are expected to embody a balance of efficiency and quality, with a strong inclination toward the latter, especially given the sensitivity of the long-term care space context.

Thus, the research presented is both academically and practically significant, as it explores an innovative collaboration between human expertise and machine intelligence. The aim is to discover how CGANs can be most effectively incorporated into the design process, ensuring that the architect’s role as the decisive authority in achieving a final design is human-centered, contextually relevant, and of high quality. By bridging the gap between novel technological approaches and traditional architectural practice, this study seeks to establish a new paradigm for the design of spaces for elderly people, where technology serves as an ally to the human-centric design philosophy.

Research Methods and Statistical Analysis of Interior Design of Long-Term Care Space Floor Plans

Research Methods and Process

This article explores the use of CGANs to assist architects in the floor plan design of long-term care spaces in retirement home buildings and in the construction of an automated design process by mimicking the design steps of human designers working in actual design projects. This research can improve the work efficiency of architects and has strong practical significance. In this study, the decision to use pixel graphics as input for generating museum floor plan designs is deeply rooted in the core architecture and advantages of Conditional Generative Adversarial Networks (CGAN).

1. CGAN models are adept at processing rich visual information and can generate images with a certain degree of detail and complexity by learning patterns from large data sets. Although pixel images may not match vector-based graphics in detail, this visual data provides a more continuous visual representation. This continuous visual information provides the model with rich context, allowing it to generate more informative preliminary design proposals. In the early stages of design, these generated pixel images can serve as a useful tool for exploring different design possibilities, helping designers quickly iterate and visualize ideas. However, we acknowledge that vector-based graphic models do have their advantages in applying design results to specific architectural designs because most designers use vector graphics for architectural design. Therefore, the focus of this study is to explore the possibility of generating intention diagrams rather than directly generating design drafts. This provides designers with a new perspective and tools to promote the development of innovative thinking in the early stages of creativity. In addition, pixel images are critical to the CGAN model training process.

2. The CGAN model utilizes Convolutional Neural Networks (CNN) as a key component, and CNNs demonstrate exceptional performance in processing pixel-graphic data. This is because CNNs can effectively recognize and extract local features in images and construct complex graphical representations in a hierarchical manner.

3. Considering the goal of the CGAN model is to autonomously generate practical and innovative design proposals, choosing pixel graphics provides an ideal foundation for achieving this objective. Pixel graphics, being the most common image format on computers, are easy to collect, modify, and process. This means that designers can conveniently use existing materials with the model as well as optimize generated designs, thereby better exploring the possibilities of design.

It can also be used as a tool for spatial research and has important practical value for the interior design of long-term care spaces. The research method involved the following five steps: data capture, data processing, model training, model evaluation, and model application (Figure 1).

Figure 1. Research flowchart of this study.

The first step in interior design is to divide the space according to the overall outline of the building. Model 1 can quickly create a variety of space division schemes to help designers get inspiration in the process. The second step is to determine the door and window layout, furniture layout and other basic facilities of each space according to the space division, and design a plan for the early design stage of the interior design project. Model 2 can further generate a plan containing the location and size information of these projects according to the spatial division scheme generated by Model 1. So far, CGAN has mimicked the conceptual process of human designers in the early design stage through Model 1 and Model 2. Next, the designer selects the best plan in the generated plan, and further improves it through the design deepening stage before it can continue in the subsequent construction stage. The stage of CGAN’s participation in assisting architectural designers in interior design is shown in the Figure 2:

Figure 2. The stage of CGAN’s participation in assisting architectural designers in interior design.

Data Capturing

This article aimed to collect a certain amount of nursing home care space data to provide training for the model to identify the characteristics of interior design of long-term care space. First, in the preliminary survey, data from 37 nursing homes were collected. These retirement home buildings vary in size, ranging from single-family buildings to complexes consisting of multiple structures, with the number of floors ranging from one to six floors. These cities also cover first- to fourth-tier cities on the Chinese mainland and are highly inclusive. The number of main functional areas in long-term care space also varies depending on the scale of the buildings, ranging from as few as 3 to as many as 11. The most important functional areas are the activity area, accommodation area, nurse station, and medical area. The final collection of retirement home building floor plans includes long-term care spaces of varying sizes and layouts. From past experience, we know that under the most ideal situation, the number of samples is 20 times the number of labels (Min, Zheng, and Chen 2023). In this study, there are six main labels for floor plans of long-term care space: walls, accommodation spaces, supporting spaces, stairwells, elevator rooms, corridors and public spaces. Some floor plans of long-term care space only have five due to a lack of elevator room labels. Therefore, the training sample size requires at least 100 ~ 120 floor plans to obtain ideal experimental results. After checking and screening, the researchers selected a total of 101 floor plans of long-term care spaces in retirement home buildings as experimental samples. These data can provide enough information for the model to explain the characteristics of interior design of long-term care spaces.

Data Processing

Following the completion of the data collection, the researchers undertook data preprocessing to enhance the efficacy of the model training outcomes and mitigate issues such as unmanageable training results. The primary techniques employed involve stepwise training of the model to regulate the generation effect exhibited by the model at each iteration. The floor plan samples that were gathered were subsequently reconfigured into three distinct categories of visual representations: floor plan profile (FPP), functional segmentation layout (FSL), and floor plan effect (FPE). One of the design elements that is represented by FPP is the scope of the floor plan. FSL reflects space division strategy in the floor plan encompasses various elements, such as walls, accommodation spaces, ancillary spaces, stairwells, elevator rooms, corridors, and public spaces. FPE denotes a greater level of precision in design elements, encompassing factors such as door and window placements, furniture arrangement, and the distribution of other essential facilities. (b) The materials mentioned above are consolidated into a 512 × 512 pixel format with a horizontal and vertical resolution of 96 dpi and a color depth of 24 bits. This consolidation aims to facilitate efficient training of the model and mitigate potential errors arising from variations in material sizes.

Model Training

The CGAN model framework necessitates the utilization of paired images as input data while adhering to the principle of incremental training of the model (the machine learning environment is described in Appendix A). Following the completion of data processing, the three categories of images (FPP, FSL, and FPE) undergo training in a sequential manner, progressing from outline to function and subsequently from function to effect. These training processes align with the two distinct models, Model 1 and Model 2. These two models encompass all the stages from the initial design phase to the ultimate outcome, thereby effectively accomplishing the experimental objective.

Model Evaluation

It is imperative to assess the training process and outcomes of the model promptly to observe the genuine impact and applicability of the model. This study employs two methodologies to assess the model: (a) It is imperative to carefully monitor the magnitude and general trajectory of the loss function at each iteration throughout the process of training the model. The loss value in CGAN is employed as a metric to evaluate the fidelity of the generated samples (DeVries et al. 2019; Salehi, Chalechale, and Taghizadeh 2020). A lower loss value indicates superior model training and a higher degree of similarity between the generated samples and the real samples. A consistent decrease in the loss value throughout the training process indicates an improvement in the model’s training. Conversely, if the loss value demonstrates a consistent increase, it suggests a deterioration in the model’s training. (b) The model is evaluated at each iteration, the test image is exported, and the presence of “image noise” is assessed through visual observation. The presence of “image noise” can decrease the discriminator’s ability to accurately distinguish between true and false judgments, leading to a decrease in the loss value. However, this can also result in the generation of a poorly rendered effect. The image within the training set can be identified as an authentic image, while the picture generated by the CGAN can be classified as a synthetic image. The term “real picture” refers to the authentic representation of a floor plan, while the term “fake picture” denotes a projected or simulated floor plan (Min, Zheng, and Chen 2023). By employing the aforementioned two evaluation methods, one can gain a fundamental understanding of the efficacy of model training, thereby facilitating the development of additional strategies to address issues encountered during the training process. These strategies include the following steps: (a) trying different optimizers, such as Adam, RMSProp, or SGD, to speed up the model’s convergence; (b) changing training parameters, such as the learning rate, batch size, and regularization parameters, to improve the model’s performance; and (c) considering changes in the network structure, such as adding or removing layers or neurons, to improve the model’s ability to generalize.

Model Application

For the CGAN model to generate viable design suggestions, input images in the form of FPP should be provided. This requirement ensures that the model’s output is grounded in relevant architectural context and spatial constraints, allowing for more tailored and applicable design recommendations. Therefore, the trained CGAN model can be applied to many aspects of the interior design of long-term care spaces, including spatial research, design assistance, and enhance the architect’s design process. At the same time, this method can also be extended to other areas of elderly care and medical systems to improve the efficiency and scientific nature of space design. It is particularly worth noting that, when generating design proposals using the CGAN model, architects are required to provide input images to the model, such as floor plan profiles or functional segmentation layouts. Based on these input conditions, the CGAN model will analyze and generate images as suggestions for the design, thereby assisting architects in exploring innovative design solutions. In addition, in order to compare the differences between architects’ designs and CGAN designs, the researchers used 10 architect-designed long-term care spaces in retirement home building floor plans as samples for model application. Through the application of the CGAN model and the final effect, it is compared with the architect’s design to evaluate the possibility of applying the model.

Material Handling

Test material samples were collected from 37 retirement home building projects on the Chinese mainland. These retirement home buildings are distributed in Beijing Province, Shandong Province, Hubei Province, the Inner Mongolia Autonomous Region, and other provinces across the Chinese mainland and cover various sizes and forms of long-term care spaces in retirement home buildings on the Chinese mainland. A total of 101 long-term care space floor plans were extracted from these retirement home buildings and standardized into images of the same size to facilitate analysis. According to the needs of the experiment, the 101 floor plans were divided into three types – floor plan profiles (FPPs), functional segmentation layouts (FSLs), and floor plan effects (FPEs) – for a total of 303 images. The FPP sample uses a grayscale image with a black color represented by the RGB (R0, G0, B0). During the processing of FSL samples, a method is employed to streamline the data by utilizing distinct colors to represent the primary elements found within the floor plan of long-term care spaces in retirement home buildings. These elements are then visually presented with blocks of different colors. In this study, six colors were used to represent elements in floor plans of long-term care spaces in retirement home buildings (Figure 3). For example, walls are green (R0, G255, B0), accommodation spaces are red (R255, G0, B0), supporting spaces are yellow (R255, G0, B0), and stairwells are sky blue (R0, G255, B255). The elevator room is dark blue (R0, G0, B255), and the corridors and public spaces are purple (R255, G0, B255). Each floor plan contains some or all of these six elements, depending on the situation, and these markers can help machine training and recognition.

Figure 3. The materials used in this study encompassed three key components: (1) the floor plan profile (FPP), (2) the functional segmentation layout (FSL), and (3) the floor plan effect (FPE).

CGAN Model

The conditional generative adversarial network (CGAN) is a special kind of generative adversarial network (GAN) that creates target data by adding conditional variables to the generator and discriminator structures (Durgadevi 2021). These conditional variables may take the form of labels, images, or any other auxiliary data so that the generated data not only learn features from the noise vector ($\mathit{Z}$) but also follow the instructions of the conditional variables, resulting in more diverse and useful generated data (Odena, Olah, and Shlens 2017; Yao et al. 2018). By providing additional information to the model, the CGAN can generate data that match the conditions, which is especially useful when the dataset is complex or a specific type of output is needed.

As shown in Figure 4, taking Model 1 as an example, the running process of the CGAN is as follows: (1) Initialization: Select a random noise vector ($\mathit{Z}$) and an input picture, such as the FPP. (2) Generator: The generator receives noise vectors and FPP images and generates sample data, such as FSL. (3) Discriminator: The discriminator evaluates the generated FSL data and the real FSL data and distinguishes between the two. (4) Training process: The weights of the generator and discriminator are updated through backpropagation and gradient descent so that the discriminator cannot distinguish between the generated FSL data and the real FSL data. (5) Iterative optimization: The above process is repeated until the generator generates high-quality FSL data; the discriminator cannot easily distinguish authentic from fake data; and finally, FSL images are generated from FPP images.

Figure 4. CGAN model framework.

Generator Architecture: The generator in our CGAN model comprises five convolutional layers. The first layer features 256 filters with a kernel size of 5 × 5 and a stride of 1, followed by layers with 128, 64, 32, and 16 filters. The filter size of each subsequent layer decreases to progressively refine the generated image details. Intermediate layers employ upsampling and batch normalization to enhance image quality. For instance, the second and third layers use 2× upsampling to gradually increase the spatial dimensions of the feature maps. The activation function in the final layer is Tanh, which normalizes the output to a range that matches the input image format, producing realistic floor plan elements.

Discriminator Architecture: The discriminator is designed with four convolutional layers. The first layer has 64 filters with a kernel size of 3 × 3, and the second layer has 128, 256, and 512 filters, each with an increased capacity to extract more intricate features from the input images. In the discriminator, a stride of 2 is used for downsampling, which effectively reduces the spatial dimensions while deepening the feature representation. The final layer is a fully connected layer that outputs a single probability score, indicating the likelihood of the input being a genuine floor plan element.

Training process: The training process involves alternating updates to the generator and discriminator. The generator starts with a random noise vector ($\mathit{Z}$) and a conditional input (e.g., an FPP image) and aims to produce floor plan elements that are indistinguishable from real ones. The discriminator, trained concurrently, learns to distinguish between the authentic and CGAN-generated floor plan elements. The learning rate for both networks is set to 0.0002 with a beta1 value in the Adam optimizer of 0.5. The process iterates until the discriminator’s accuracy in distinguishing real from fake images plateaus, indicating the generator’s ability to produce high-quality, realistic floor plan elements.

G_GAN: This is the primary generator component within the CGAN framework and is responsible for creating synthetic images. In our model, G_GAN takes the random noise vector ($\mathit{Z}$) and conditional inputs (such as FPP images) and synthesizes preliminary floor space layouts (FSLs). This component is essential for initiating the generative process and laying the groundwork for more refined interior space design elements.

G_GAN_Feat: Serving as a feature extraction module within the generator, G_GAN_Feat is tasked with interpreting and incorporating specific architectural features from the input data. The conditional inputs, such as labels indicating room types or design requirements, are analyzed, and the generator is guided to produce outputs that adhere to these specified conditions. This feature extraction is vital for ensuring that the generated designs are not only innovative but also contextually relevant.

G_VGG: Employing the renowned VGG network architecture, this component enhances the generator’s ability to extract complex features and refine the image synthesis process. The G_VGG module aids in producing more detailed and accurate interior space designs by understanding deeper patterns and structures within the input data. Its role is particularly significant in achieving a high level of detail and realism in the generated floor plan layouts.

D_real and D_fake: These are two critical discriminator components. D_real evaluates the authenticity of actual floor plan images, serving as a benchmark for real interior space design. D_fake, on the other hand, assesses the generator’s output. It determines how closely the synthetic designs mimic real floor plans, challenging the generator to improve its output iteratively. The interaction between D_real and D_fake is fundamental to the adversarial training process, pushing the entire model toward producing more realistic and accurate interior space designs.

In this study, by providing CGANs with relevant conditions for long-term care spaces, such as FPP or FSL, CGANs can generate FSLs or FPE that meet specific conditions. This approach not only improves design efficiency but also increases design diversity, providing architects with a powerful tool to generate design solutions tailored to the specific needs of older adults. In addition, the results generated by the CGAN can be used as an initial draft of interior space design, providing a basis for further design decisions and modifications, accelerating the design process, and possibly improving the age-appropriate quality of interior space.

Model Training

Before model training, basic settings and adjustments should be made to the model so that it can reflect refined and professional processing methods according to the characteristics of the floor plan design of long-term care spaces in a retirement home building. Specifically:

1. The batch size is set to 1, which ensures that the model can focus on processing each design sample one by one. This is especially important when learning complex and detailed interior space designs.

2. The beta1 parameter of the Adam optimizer was set to 0.5. This is an empirical value derived from repeated testing in this study, which can effectively balance the exponential decay rate of first-order moment estimation. Therefore, the training stability and convergence speed improve, especially during the training process of the CGAN.

3. The data type is specified as 32 bits, which not only reduces the consumption of computing resources but also ensures sufficient accuracy, making it particularly well suited for processing the complex and voluminous data typically involved in high-dimensional architectural design.

4. The display frequency is set to display the training results once every 100 batches. This allows researchers to regularly monitor the learning progress and performance of the model and adjust the training strategy in a timely manner.

5. The size of the training image material is set to 512 × 512 pixels, ensuring that the details of each design can be clearly displayed when visualizing the generated interior space design. These parameter settings for these models balance efficiency and performance, focusing on improving the model’s ability to handle the complexity and detail of interior space design. The following section provides a detailed analysis of the process and results of model training after these settings.

Training Process

When studying machine learning models, analyzing the model’s loss value is crucial to understanding and optimizing the model’s performance. The loss value reflects the difference between the predictions and the actual target values and the actual results and is an important indicator for evaluating the learning effect of the model. In the context of the floor plan design of long-term care spaces in retirement home buildings, this indicator is particularly important because it directly affects whether the design solution generated by the model meets the actual needs and expected effects.

The loss value indicators include G_GAN, G_GAN_Feat, G_VGG, D_real, and D_fake. These indicators reflect the different loss values of the generator (G) and the discriminator (D) in the generative adversarial network (GAN) and are crucial for evaluating model performance and training effects. Among them, G_GAN represents the loss value of the generator in the GAN architecture. A lower G_GAN indicates that the data generated by the generator are more able to deceive the discriminator; that is, the generated architectural interior space design plan is more realistic. G_GAN_Feat is the generator feature matching loss, which measures whether the features of the generated data are close to the features of the real data. G_VGG is the perceptual loss based on the VGG network and is used to measure the visual difference between the generated design drawings and the real design drawings. D_real and D_fake represent the loss values of the discriminator on real data and generated data, respectively. A lower D_real and a higher D_fake mean that the discriminator can effectively distinguish between real and generated designs.

As shown in Figure 5 below, during the training process of Model 1, the researchers observed that the maximum value of G_GAN was 0.8, the minimum value was 0.3, and the average value was approximately 0.5. The maximum value of G_GAN_Feat is 1.2, the minimum value is 0.4, and the average value is approximately 0.7; the maximum value of G_VGG is 1.3, the minimum value is 0.2, and the average value is approximately 0.6. The D_real and D_fake values fluctuate between 0.3 and 0.6 and between 0.4 and 0.7, respectively. This shows that Model 1 performs well in generating realistic design solutions, but there is still room for improvement in feature matching and visual perception. In contrast, Model 2 is more stable in terms of the performances of G_GAN, G_GAN_Feat, and G_VGG, and its loss value fluctuates less, indicating better learning and adaptability. For example, G_GAN has a maximum value of 0.7, a minimum value of 0.2, and an average value of approximately 0.4. The maximum value of G_GAN_Feat is 1.0, the minimum value is 0.3, and the average value is approximately 0.6; the maximum value of G_VGG is 1.1, the minimum value is 0.1, and the average value is approximately 0.5.

Figure 5. Line chart of loss values during model training: (1) loss of model 1; (2) loss of model 2.

A specific analysis of these indicators reveals that Model 2 has better performance in generating FPE and can better capture and imitate real-world interior space features. However, there is room for improvement in the loss values of the two models in terms of feature matching and visual perception. To further test the performance of the model, the following uses the model with different iterations to generate images and uses images to explain the model training process more intuitively.

Figure 6 shows the progress and changes in the building floor plans generated by Model 1 and Model 2 under different training generations. Compared to the actual floor plan, Model 1’s image at the fifth epoch has a messier color distribution, indicating that the model has not yet learned enough features to accurately reproduce architectural interior space elements. As the training generations increase to the 100th and 200th epochs, the elements in the generated images start to become increasingly clearer and more orderly. Especially in the 200th epoch, we can already see clearer spatial divisions and layouts of architectural interior space elements. For Model 2, judging from the generated image of the 5th epoch, although the structure is quite different from that of the real image, the distribution of its elements and the definition of color blocks are clearer than those of the initial generated image of Model 1. By the 100th and 200th epochs, the images generated by Model 2 show higher structural consistency and approximation to real images. Especially at the 300th epoch, the spatial layout of the generated image is more similar to that of the real image, demonstrating that the ability of the model to learn architectural interior space design elements and spatial relationships is constantly improving. Overall, as the number of training generations increases, the building floor plans generated by the two models become increasingly closer to the real floor plan. This reflects the model’s progressive improvement in identifying and simulating key features of the interior design of long-term care spaces. This advancement is crucial for machine learning-assisted interior space design, as it suggests the potential for models to assist in generating design solutions that align with practical needs, serving as a tool to help designers obtain diversified design solutions for reference in the conceptual design stage.

Figure 6. Test images of the model accuracy during the iteration process.

Model Testing

After the model training is completed, model testing is an important step for verifying the model’s generalization ability, which is intended to confirm whether the model can accurately learn from the training data and generate new instances. As shown in Figure 7 below, we use Model 1 to test Samples 1 to 3. The input graphics in Sample 1 are complex and contain multiple indented corners and convex parts. Although the image generated by Model 1 retains the main structural features, it still has distortions in some details, such as the processing of corners and the division of internal space. The input of sample 2 is relatively simple, and the image generated by Model 1 in this case is closer to the input, demonstrating the efficiency of the model in processing simpler structures. The input graph of Sample 3 is in the shape of an “E,” and the results generated by Model 1 maintain the overall structure. The misalignment and color confusion indicate the limitations of the model in processing graphs with complex internal structures.

Figure 7. The training samples were used to test the final model.

Model 2 was used to test samples 4 to 6, which were further generated based on samples 1 to 3. Model 2 provides more advanced detail and spatial reconstruction capabilities for these samples. Sample 4 shows that the model has a more reasonable layout of the internal space while maintaining the overall structural frame. The results of Sample 5 show the model’s clearer delineation of the boundaries of different functional areas in the floor plan. Sample 6 reveals how the model can effectively infer and generate accurate spatial layouts when processing more regular input graphics.

Overall, the images generated by Model 1 and Model 2 reveal their unique capabilities in identifying and generating interior design plans of long-term care spaces. Model 1 is suitable for handling designs with simple structures, while Model 2 is more effective at handling more complex and detailed designs.

Multischeme Generation Effect Test

Researchers can change the generative process to obtain different design solutions by changing the noisy input vectors of a CGAN. This technique works by feeding different random noise vectors into the network, which serve as an input to the generator, causing variation in the generated output. This variety is particularly useful for interior space design because it allows designers to choose from a range of possibilities and create solutions that meet specific functional and aesthetic requirements.

As shown in Figure 8 below, Model 1 and Model 2 generated eight different design solutions for the same input image. The design proposals generated by Model 1 explore a variety of architectural spaces, including different layouts of accommodation spaces, supporting spaces, and public areas. These schemes provide a different combination of spatial circulation and functional areas while maintaining the integrity of the structural core. For example, some schemes emphasize the continuity of accommodation spaces, while others provide greater openness to common areas.

Figure 8. A test of the model’s ability to generate multiple design solutions.

Model 2 shows more refined spatial layout processing during the generation process. The generated floor plan shows the optimization of corridors and public spaces while maintaining the structural foundation, providing smoother circulation and reasonable space allocation. The Model 2 solution is more refined in detail, such as the layout of the stairwell and elevator room. Given its advanced features, Model 2 has the potential to understand and apply building codes, which warrants further investigation.

According to the generated results, both models can create design solutions that conform to the actual architectural interior space logic based on the input building layout. Model 1’s generated scheme was diverse in spatial layout, while Model 2 showed a deep understanding of interior space detail and practicality. This multi-scheme generation method not only enhances the innovation and personalization of design but also provides a wider range of options for interior space design decisions. Through these diverse designs generated by CGANs, architects and planners can evaluate and select interior space options that best suit specific needs and site conditions, but these optional design schemes still need to be further improved in the following design stages.

Discussion

Model Application

The model’s ability to work with new samples is tested by performing experiments with six newly made FPP images, as shown in Figure 9. Initially, the six recently produced FPP images were generated through the utilization of Model 1. Subsequently, the images generated by Model 1 are employed as the input for Model 2, culminating in the generation of the ultimate image by Model 2. Based on the comprehensive findings of the study, we have concluded that the CGAN model has potential to participate in the design process, but there are still some shortcomings.

Figure 9. The new FPP of long-term care spaces in retirement home buildings was utilized to test models 1 and 2.

The Indispensable Role of Architects

The interplay between Models 1 and 2 within the interior space design process underscores a symbiotic relationship between machine-generated creativity and the indispensable expertise of architects. The challenges and imperfections observed in the outputs from Model 1, which in turn influence the effectiveness of Model 2, are not seen as drawbacks but rather integral components of a collaborative design framework. These instances of errors and the necessity for manual refinement emphasize the ongoing need for the architect’s creative input, not as a mere corrective force but as an essential part of the iterative design process. This dynamic ensures that the initial designs generated by machine intelligence serve as a foundation upon which architects can build, refine, and optimize, leveraging their unique skills and perspectives to elevate the overall design quality. The relationship between the models and architects highlights a process where technological innovation and human creativity converge, fostering a design approach that is both innovative and grounded in practical expertise.

Strategic Design Solutions for Elderly Well-Being

The results generated from Test 1 to Test 6 reveal a strategic layout where living spaces are concentrated and support areas are dispersed, with stairwells and elevators situated away from living quarters. This design minimizes disturbances for residents and promotes easier interaction, helping to alleviate loneliness and anxiety among the elderly. Additionally, the design of corridors and public spaces in Test 2 and Test 4 adopts a more open and accessible approach, catering to the elderly’s needs for leisure and socialization. This alignment with contemporary design principles not only meets the physical and psychological needs of residents but also showcases the learning potential of this model in creating supportive environments for the well-being of elderly individuals in retirement homes.

Impact of FPP Regularity

Although machine learning demonstrates potential in aiding the design of long-term care spaces, further analysis of the results reveals the pivotal role of FPP regularity in producing feasible results. Specifically, outputs from more regular FPP images, as seen in Test 1 and Test 3, yield more logical and architecturally sound designs suitable for the needs of the elderly and their caregivers. The outcomes of Test 2 and Test 6, conversely, were perplexing and deviated from fundamental architectural interior space principles due to their reliance on highly distorted FPP images. As a result, they had little use in practical interior space or as a source of design inspiration. Additionally, the analysis has uncovered specific shortcomings across various tests. The findings from Test 5 and Test 6 indicate a lack of comprehensive linkages between the halls, public spaces, and other locations. This could make it harder for users to find their way around and make the space more useful, which makes it hard to come up with a workable plan for the early ideation phase. Similarly, the results from Test 4 and Test 5 are criticized for having an insufficient proportion of supporting spaces. Their stairwell and elevator plans are also deemed inadequate, not fully accommodating the needs and comfort of elderly residents and their caregivers. Notably, the result from Test 3, despite being more reasonable overall, places stairwells too close to living areas, potentially disturbing the daily lives and sleep quality of residents due to noise.

Analysis of Challenges in Application of CGAN Model

In summary, the floor plan design of long-term care spaces in retirement home buildings generated by the CGAN model exhibits functional integrity, providing a valuable framework for further design and development. This approach provides architects with more choices and greatly improves their design efficiency. However, there are many problems in practical applications, and the generated results are far from expected. The possible reasons are as follows: (1) The newly drawn FPP images are very different from the sample images in the training set, preventing the model from accurately generating the expected results. However, the generalizability of this model needs to be further improved. (2) The number of samples needs to be supplemented. This study used a total of 101 long-term care spaces in retirement home building floor plans. There are as many as six labels, which may not meet the complex learning needs of the model. However, follow-up research needs to increase the sample size and optimize the model to improve its predictive ability and accuracy. In addition, in practical applications, the physical environment of different regions, such as temperature and lighting, should also be considered, and the specific needs of elderly people and their caregivers should be considered to further optimize the interior design of long-term care spaces. Therefore, models cannot completely replace architects in completing all design work, but they can still provide architects with reference designs for different solutions at the conceptual design stage.

Comparative Analysis of Differences with architects’ Designs

To evaluate the performance of this CGAN model and its similarities and differences with architects’ designs, the researchers selected four actual interior design projects of long-term care spaces projects for comparison. The FPP image is extracted from it as the input of Model 1, the image generated by Model 1 is used as the input of Model 2, and the floor plan of long-term care spaces is generated (Figure 10). To enhance the effectiveness of Model 2 and prevent any issues that may arise from its use, the step “Fixing Model 2” was included. Based on Model 1, the photographs created by the model are manually adjusted by performing simple modifications. These corrections involve lowering picture noise and enhancing the clarity of color block edges. The designer screens the images of FPE generated by the machine and fix the selected FPE which is more valuable, ensuring its effectiveness. In the actual design projects, this step belongs to the last step of the conceptual design stage. Figure 10 shows the interior design projects for long-term care spaces (Model 2) generated by the CGAN model and its subsequent manually corrected version (Model 2_fixed). The conversion from Model 2 to Model 2_fixed involved a critical editing step that represented the transformation from the initial machine-generated design to a more rational and practical floor plan. Although the model can generate a rough layout of the long-term care spaces based on the training data in Model 2, the results often lack the precision and accuracy required in the actual architectural interior space design process. Manual editing was necessary to resolve these issues, which included resizing spaces, optimizing corridor layouts, and ensuring the proper placement of elevator shafts. This method of human-machine collaboration transforms the preliminary output of Model 2 into Model 2_fixed, better aligning with architectural design specifications. This manual editing step is not only necessary but also extremely useful. It makes up for the shortcomings of the current CGAN model in generating indoor floor plans, bringing the final design results closer to what the actual indoor space design needs.

Figure 10. Comparative analysis of differences with architects’ designs in floor plan design of long-term care spaces.

To evaluate the effect of the model, further improve the model performance, and analyze the applicability and limitations of the model, the researchers compared the final results generated by the CGAN model with the actual project of interior design of long-term care spaces. Additionally, we constructed a 3D model for display and comparison, using six colors to distinguish the elements of the long-term care space plan in the retired home buildings (Figure 11). For example, the walls are green (R11, G179, B11), the accommodation spaces are red (R205, G32, B36), the supporting spaces are yellow (R255, 167, B21), and the stairwells are sky blue (R54, G152, B215). The elevator room is dark blue (R41, G1, B196), and the corridors and public spaces are purple (R118, G9, B138). This corresponds to the design deepening stage in the actual design projects which requires the designer to carry out aesthetic and functional optimization.

Figure 11. A 3D model comparison based on architectural design and CGAN model design results in the floor plan design of long-term care spaces.

The comparison results are as follows:

1. There are certain differences between CGAN model design and manual design, and manual design is better than model design. In the four projects, the main purpose of CGAN model design is to arrange the partitions in a fixed space. The partitions generated by straight edges are more regular, while the partitions generated by bevel or irregular edges are chaotic or even impossible to form. Manually designed partitions take more into account the rational use of space, and the planning of corridors and public spaces is also more purposeful and functional.

2. The CGAN model, although not as refined as expert architects, is a valuable tool for beginners because it offers diverse design possibilities and aids spatial understanding. It is crucial, however, for novices to deeply engage with design principles rather than solely relying on the model. For experienced architects, the model can expedite proposal stages and stimulate creative alternatives. Its effectiveness may vary, so empirical studies are needed to determine its true impact on the quality of the design scheme.

3. The intrinsic strength of the CGAN model lies in its efficiency and expediency, which enable the rapid generation of a multitude of design plans. These attributes could greatly benefit the initial stages of design, providing architects with varied spatial configurations for exploration and refinement. While the model offers promising potential, it is not yet at the stage where it can substitute for the nuanced understanding and creative judgment of architects. The progressive development of such machine learning models, however, hints at an exciting future direction for interior space design, where the blend of human expertise and machine learning innovation could redefine the traditional design process.

Analysis of Public Questionnaire Scoring Results

This study seeks to critically examine the cooperative relationship between designers and machines in the creation of interior design of long-term care spaces. To achieve this goal, we conducted quantitative analysis methods with a questionnaire in which 413 design practitioners responded. In the survey, participants rated their satisfaction with all the spatial design plans using a 5-point Likert scale, with 1 indicating the lowest satisfaction and 5 indicating the highest satisfaction. The Cronbach’s alpha coefficient, a statistical measure that assesses the reliability of a set of surveys by quantifying how closely related a set of items is to a group, was used to evaluate the consistency of the questionnaires in the evaluation of long-term care spaces in interior design of long-term care space. The Cronbach’s alpha coefficient of the questionnaire was 0.951 (Table 1), which far exceeded the conventional standard of 0.7 (Duhachek, Coughlan, and Iacobucci 2005). From a statistical perspective, this high degree of internal consistency indicated that the reliability of the questionnaire was high. The results of the reliability analysis ensure the stability and consistency of subsequent studies when analyzing the data.

Table 1. Measurement for the questionnaire.

Number of items	Sample size	Cronbach’s Alpha
20	413	0.951

Source: Data sourced from a questionnaire survey conducted by the research group

Based on the reliability analysis, the researchers further conducted a validity analysis to evaluate whether the questionnaire items could accurately reflect the relevance of long-term care spaces in retirement home building design. The Kaiser – Meyer – Olkin (KMO) test and Bartlett’s sphericity test were used to assess the suitability of the questionnaire data for factor analysis. The KMO test measures the sampling adequacy for factor analysis, with a value above 0.6 indicating that the sample is suitable for factor analysis. Bartlett’s sphericity test was used to check for significant correlations among the observed variables, confirming the applicability of the questionnaire factors (Shrestha 2021). In this study, the KMO value was 0.875 (Table 2), which is higher than the conventional standard of 0.6, indicating that the sample is very suitable for factor analysis (Zeynivandnezhad, F., et al. 2019). The results of Bartlett’s sphericity test showed that there was a significant correlation between the observed variables (p < .001), which confirmed the applicability of the questionnaire factors. The results of the validity analysis not only verified the suitability of the questionnaire data but also further enhanced the robustness of the validity analysis results of this study.

Table 2. KMO and bartlett’s test for the questionnaire.

Kaiser‒Meyer‒Olkin Measure of Sampling Adequacy(KMO)	Bartlett’s Test of Sphericiry
	Approx. Chi-Square	df	Sig
0.875	5544.760	413	0.000

Source: Data sourced from a questionnaire survey conducted by the research group

In the spatial design comparative analysis, the researchers compared two different design options. A practicing architect created Design Plan A, which maintains the long-term care spaces’ conventional layout and functional arrangement. Design Plan B is the result of the designer’s optimization of the scheme generated by the CGAN model proposed in this article, showing the results of the new design method of collaboration between the machine and the designer. The images used for evaluation in the questionnaire are not directly generated by the CGAN model. Instead, these images are the output of manual corrections made by the architect using CAD modeling tools (Model 2 - Fixed). The purpose of this step is to convert the raw output of the CGAN model into more accurate and detailed architectural interior design drawings so as to more accurately reflect the standards and details in actual interior space design. Therefore, individuals involved in the evaluation were asked to compare the manually corrected CGAN drawings with the architect’s hand drawings. It should be noted that the evaluation results may reflect the impact of manual corrections and not just the accuracy of CGAN model generation capabilities. Although this approach improves the usability of the drawings to a certain extent, it also shows that, under the current technical level, the drawings generated by the CGAN model still need to be further processed by professional architects before they can meet the standards for practical applications. As shown in Table 3, although there are differences in the mean scores between the two schemes, these differences are not significant, which shows that in practical applications, both traditional design methods and cooperative design methods with machine learning models each have their own advantages. However, through comparative analysis, it was found that Design Plan B had unique advantages in terms of innovation and flexibility. These advantages may stem from the ability of the CGAN algorithm to process complex data and recognize patterns.

Table 3. Item statistics.

Shape Group	Design Plan	Mean	Standard Deviation	Responses
Comparison of “Z” shaped type layout in long-term care space spatial layout	A	3.389831	1.155650	413
	B	3.331719	1.181938	413
Comparison of “H” shaped type layout in long-term care space spatial layout	A	3.726392	1.068165	413
	B	3.685230	1.037033	413
Comparison of “I” shaped type layout in long-term care space spatial layout	A	3.840194	1.049101	413
	B	3.832930	0.984679	413
Comparison of “U” shaped type layout in long-term care space spatial layout	A	3.733656	1.073401	413
	B	3.786925	1.055785	413
Comparison of leaf-shaped type layout in long-term care space spatial layout	A	3.740920	1.193915	413
	B	3.731235	1.118215	413
Comparison of spindle shape type layout in long-term care space spatial layout	A	3.576271	1.117868	413
	B	3.583535	1.084274	413
Comparison of elliptical shape type layout in long-term care space spatial layout	A	3.915254	0.987831	413
	B	3.769976	1.015831	413
Comparison of “V” shaped type layout in long-term care space spatial layout	A	3.581114	1.119696	413
	B	3.539952	1.088990	413
Comparison of “J” shaped type layout in long-term care space spatial layout	A	3.421308	1.160319	413
	B	3.433414	1.131434	413
Comparison of “E” shaped type layout in long-term care space spatial layout	A	3.469734	1.164682	413
	B	3.539952	1.130545	413

Source: Data sourced from a questionnaire survey conducted by the research group

Through further comparison, it was found that among the design plans of U-shaped long-term care spaces and E-shaped long-term care spaces, participants rated Design Plan B with a significantly higher satisfaction score than did Design Plan A. By comparing the distribution of spatial functional areas, the researchers speculated that the reason for the high degree of satisfaction is that the supporting space occupies a large proportion and that the elevator room, stairwell, and accommodation space are far apart. Among these spaces, U-shaped long-term care spaces also have several advantages: the design of multiple elevator rooms is helpful for elderly people with limited mobility (Figure 12). However, the satisfaction with Design Plan A is much greater than that with Design Plan B. In the design of elliptical-shaped long-term care spaces, the proportions of supporting spaces, corridors, and public spaces in Plan B are too low, and the living space is too dense, which makes people feel depressed. The distance between the stairwell and the living area is too great, and the design of two elevator rooms in the middle of the corridor renders the stairwell design useless. Through comparative analysis, the researchers found that Design Plan B had unique advantages in terms of innovation and flexibility. These advantages may stem from the ability of the CGAN algorithm to process complex data and perform pattern recognition. However, there are still shortcomings in humanization and rationalization, and professional architectural design experts are still required to pick up feasible plans from a large number of generative plans and continuously improve them, reflecting the key role of designers in supporting design decisions and the efficiency of CGANs in providing conceptual proposals.

Figure 12. Score comparison between the architect design and CGAN model design.

In addition, the researchers also specifically evaluated the two options by practitioners in the architectural design and engineering industries. These professionals have shown a high interest in Design Plan B. Further interviews revealed that they believe that the application of CGAN technology in space design is highly innovative and may have an important impact on the future interior design of long-term care spaces. Their evaluation highlights the benefits of design and machine cooperation, particularly in terms of the efficiency and flexibility of designs, which may contribute to the development of future interior design of long-term care spaces.

Conclusions

This article aimed to explore the interior design of long-term care spaces in retirement home buildings assisted by CGANs. With the continuous increase in the number of elderly people in China, people are paying increasing attention to elderly people, which are regarded as vulnerable groups. This study uses CGAN to build a generative tool that mimics the design process of human designers designing long-term care space plans in retirement home buildings. CGAN is not directly involved in the design process. It still requires humane designers to provide the prototypes of architecture and screen the generation scheme. As a catalyst for stimulating innovation and creative thinking, these generated materials provide references for designers and are further improved by designers before becoming effective design solutions. The results of this study show the following: (1) For the interior design of long-term care spaces, the CGAN model has strong understanding and calculation capabilities. The zoning layout of long-term care spaces in retirement home buildings was successfully established, and the results showed that the CGAN model can serve as a valuable tool in this field of research. (2) Despite these differences, there are still many similarities between the designs created by CGANs and manual designs, demonstrating that CGANs have a strong capacity for learning and can produce better results when the sample size is large enough. (3) Both the model training results and the results can be exported from CGAN into easy-to-understand images, helping researchers evaluate and analyze them. (4) The CGAN can apply learned content to new samples and generate rich floor plans with good results. This helps architects create more detailed plans based on the model, which greatly improves work efficiency. Moreover, this approach can also explore additional space design possibilities and, to a certain extent, correct architects’ subjective assumptions. (5) Through an in-depth comparative analysis of Plan A (the plan designed by the architect) and Plan B (the plan designed by the CGAN model) through a satisfaction questionnaire, the researchers not only revealed the differences between traditional design methods and artificial intelligence-assisted design methods but also discussed how to balance functionality and innovation in the interior design of long-term care spaces. Traditional design methods rely on the professional knowledge and experience of architects, which play an irreplaceable role in ensuring the functionality and adaptability of space layouts. However, with the development of technology, the introduction of artificial intelligence technologies such as CGANs has brought new perspectives and possibilities to the design process, especially in terms of innovation and flexibility, showing significant advantages.

Although the CGAN model shows certain technical potential, its effectiveness in actual architectural design applications is still very limited. For instance, the CGAN model generates pixel images with obvious deficiencies, making it unsuitable for direct application in the designer’s workflow. It is more suitable for application in the conceptual stage of early program design. We identified this as a crucial finding during the research process, and future work must address this issue. The researchers will also need to continue to overcome these shortcomings and utilize CGAN more meaningfully in architectural design. In addition, to address these challenges, future research can explore how to improve the resolution and quality of images generated by CGAN. Alternatively, researchers could devise innovative techniques to enhance the efficiency of converting the generated pixel images into vector formats, thereby enhancing their integration with the standard architectural design process. These improvements will open new possibilities for CGAN’s application in the field of architectural design and may increase its practicality and relevance in practical work. At the same time, there are also things to consider: (1) The CGAN cannot adapt to local conditions and generate interior design plan of long-term care spaces that are most suitable for the region. Due to the differences in construction environments caused by regional differences, factors such as sunshine and climate have not been taken into account. (2) CGAN training requires a large amount of real floor plan design data for training. When the sample size is insufficient, CGAN may find it difficult to capture its characteristics, and the model will produce many uncertain factors. However, it is still unclear how much sample size is required to be accurate enough. In future research, it will be necessary to refine and improve the CGAN model, striving to further enhance its image generation quality and accuracy. (3) The CGAN can generate only floor plan designs that meet certain conditions, and the generated designs still depend on the sample set provided by the researcher. This may result in the generated floor plan design lacking creativity and novelty and failing to keep pace with the latest and most trendy designs. (4) There may be certain deviations in the generated floor plan design. In actual application, the cooperation of professional architects is required to complete the design project, which does not completely liberalize productivity.

Future research can further explore the application of CGANs in different types of space design and how to optimize the CGAN algorithm to better adapt to specific design needs. In the future, the following steps could be taken to improve and develop: (1) use more data and more refined models to train generative adversarial networks, which would improve the quality and accuracy of the generated floor plan designs; and (2) develop additional evaluation methods, including human-computer interaction, laboratory assessment, and real-world appraisal, to more accurately measure the quality and accuracy of the generated floor plan designs. (3) Exploring methods that enable architects to control and navigate the latent space of the CGAN model. This could involve manual navigation of the input space, allowing for fine-tuning of the design by the user. Understanding and manipulating the latent space could provide architects with a more intuitive and direct way to influence the generated design outcomes, leading to more customized and diversified floor plan designs. In addition to the interior design of long-term care spaces in retirement home buildings, this method can also be expanded and applied in other similar fields, such as the floor plan design of health care spaces, the floor plan design of clinic spaces, the floor plan design of hospital spaces, and the private floor plan design of nursing spaces.

In addition, research should consider how to combine traditional design methods with artificial intelligence technology to leverage the advantages of both methods and create space design solutions that are both practical and innovative. Researchers also need to pay attention to the transparency and interpretability of artificial intelligence decisions in the design process to ensure that the design results can be widely recognized by the industry. In the future, when designing long-term care spaces, researchers recommend that designers consider an approach that blends traditional expertise and artificial intelligence technology. This approach not only improves the efficiency and quality of the design but also provides a more comfortable and beautiful living environment for nursing home residents while meeting functional needs. Through this integrated approach, the challenges brought by the aging population can be better addressed, and a greater quality of life can be provided for elderly individuals.

Author Contributions

Yanyu Li wrote the first draft, drew preliminary illustrations, and designed the research questionnaire; Huanhuan Chen revised the drawings, drew the 3D model during the experiment, and analyzed the background of the research. Jingyi Mao was deeply involved in the early stages of the survey and the digital processing of samples. Yile Chen and Liang Zheng revised the first draft, constructed the entire research idea, and translated and revised the English. Liang Zheng conducted the training of the machine learning model. Huanhuan Chen, Junjie Yu, and Lulu He conducted extensive questionnaire surveys and statistics. Junjie Yu sorted out the literature review of the study and wrote the analysis of the questionnaire results. Lina Yan provided technical assistance throughout the research process. Jingyi Mao, Yile Chen, and Liang Zheng completed the drawing and revision of the paper’s illustrations, primarily producing three revised versions of the article. All authors have read and agreed to the published version of the manuscript.

Disclosure Statement

No potential conflict of interest was reported by the author(s).

Data Availability Statement

The original code of the program cannot be released yet because our program is being used in other research. The training data set can be downloaded here: https://data.mendeley.com/datasets/6j9fpjc7yd/1

Supplementary Data

Supplemental data for this article can be accessed online at https://doi.org/10.1080/08839514.2024.2354090

Additional information

Funding

This paper is the 2023 Qingyuan City philosophy and social science planning project, Research on the construction of community service system for rural enjoyment of elderly life in Qingyuan City under the background of rural revitalization（Project Number: QYSK2023129）and Guangdong Provincial Department of Education’s key scientific research platforms and projects for general universities in 2023: Guangdong, Hong Kong, and Macao Cultural Heritage Protection and Innovation Design Team (Funding Project Number: 2023WCXTD042).

Appendix A:

Machine learning environment configuration

Machine learning environment configuration: the operating system was Windows 11 (X64), the Cuda version was 11.5, the deep learning framework was Pytorch, the graphics card was GeForce GTX 3070 (16 G), and the processor was AMD Ryzen 9 5900HX (3.30 GHz).