Physical information neural network combined with the symbolic regression: A machine learning method for prediction of dielectric constants in organic liquids and water mixtures

Although numerous machine learning models have successfully predicted material properties with favorable outcomes, the absence of a feasible and efficient analytics platform for quickly introducing molecular descriptors with interpretable potential has made it costly and required expertise from a materials field specialist. We propose a physical information neural network combined with the symbolic regression method (PINN-SR) to bridge this gap and enhance data analysis efficiency. We developed a Generative Pre-trained Transformer specialized platform (GPTS) named the Dielectric Constant Predictor, based on custom ChatGPT versions specifically tailored to present and analyze the dielectric constant of organic liquids and mixtures with water results. Our findings indicate that PINN-SR not only demonstrates robustness in discovering correct physically meaningful symbolic expressions from data but also shows integrating domain knowledge can significantly enhance performance, achieving an R² value of 0.994. Moreover, this work can accelerate mechanism exploration and provide generalized, convenient models for liquid material science.