Predicting the melting point of liquid crystals with directed message passing neural networks

ABSTRACT

Liquid crystal (LC) phases must exist in an appropriate temperature range for practical applications and thus the melting point (MP) is a critical design target of LCs. In this work, we have evaluated the performance of directed message passing neural networks trained on a large database of organic molecules (27,000+) in the prediction of a structurally diverse set of 780 LC and LC-like molecules including cyano-, azo-, ester, cyclohexyl- and halogen-containing compounds. Our results show that the model can provide accurate MP predictions for LC and LC-like molecules with an overall Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) of 23°C and 30°C, respectively. Our model provides similar levels of performance with MAEs of 20°C, 20°C, 23°C, 25°C and 24°C for azo, cyclohexyl, ester, halogen and nitrile compounds, respectively, and is most accurate for molecules with MP between 50 and 150°C (MAE = 21°C). We have developed an online tool called LCMelt (v1.0) (lcmelt.streamlit.app) where our model can be used to predict the MP of potential LC candidates before synthesis free of charge.

Graphical abstract

KEYWORDS:

• Machine learning
• melting point prediction
• liquid crystals
• Neural Network

Acknowledgments

A.S would like to acknowledge the financial support of the University of Alabama Graduate School as a Graduate Council Fellow. A.S and T.S would also like to thank the University of Alabama and the Office of Information Technology for providing high-performance computing resources and support that has contributed to these research results. This work was also made possible in part by a grant of high-performance computing resources and technical support from the Alabama Supercomputer Authority.

Disclosure statement

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