Benchmarking machine learning and molecular dynamics methods for the prediction of the thermophysical properties of organic liquids

Abstract

Accurate prediction of thermophysical properties is crucial for the rational design of lubricants and cooling fluids. We compile a database of approximately 4700 critically evaluated reference values for the density, viscosity, and thermal conductivity of 835 unique organic molecules in the C10–C50 range. We use this database to train and benchmark a range of machine learning (ML) architectures and molecular representations. XGBoost models trained on molecular descriptors are consistently among the most accurate across all three target properties. Out-of-distribution tests show that descriptor-space distance and the position of the predicted value within the training range identify higher-risk regions at the population level, rather than ordering the errors of individual molecules. Equilibrium molecular dynamics (EMD) simulations of eight representative molecules reproduce density accurately but show systematic, chemistry-dependent force-field errors for the viscosity and thermal conductivity. At an inference cost orders of magnitude lower than EMD simulations, the ML models are well suited to high-throughput screening of new candidate molecules for lubricants and other formulated products.