Synergistic Modeling\nof Liquid Properties: Integrating\nNeural Network-Derived Molecular Features with Modified Kernel Models

A significant challenge in applying\nmachine learning\nto computational\nchemistry, particularly considering the growing complexity of contemporary\nmachine learning models, is the scarcity of available experimental\ndata. To address this issue, we introduce an approach that derives\nmolecular features from an intricate neural network-based model and\napplies them to a simpler conventional machine learning model that\nis robust to overfitting. This method can be applied to predict various\nproperties of a liquid system, including viscosity or surface tension,\nbased on molecular features drawn from the <i>ab initio</i> calculated free energy of solvation. Furthermore, we propose a modified\nkernel model that includes Arrhenius temperature dependence to incorporate\ntheoretical principles and diminish extreme nonlinearity in the model.\nThe modified kernel model demonstrated significant improvements in\ncertain scenarios and possible extensions to various theoretical concepts\nof molecular systems.