A hybrid molecular dynamics–machine learning framework for boiling point estimation in aromatic fluids

Precise estimation of boiling points in organic fluids is critical for designing efficient and safe thermal systems. This study presents a hybrid molecular dynamic (MD)–machine learning (ML) framework for boiling point estimation in two representative aromatic fluids: biphenyl (C12H10) and diphenyl ether (C12H10O). Two force fields, OPLS-AA and COMPASS, were tested in equilibrium MD simulations. OPLS-AA produced density predictions with a relative error below 2 % compared to experimental values, while COMPASS showed reduced accuracy at elevated temperatures. Boiling point was estimated using a density threshold method (yielding 525.66 K) and a thermodynamically rigorous inflection-point method (508.18 K), revealing ∼3.3 % deviation between boiling onset and completion. MD data were used to train and evaluate three regression models—Nearest Neighbours Regression (NNR), Neural Network (NN), and Support Vector Regression (SVR). The NNR model achieved the best match with MD data, predicting a boiling point of 524.97 K and density of 0.064 g/cm3. The NN model accurately estimated boiling temperature (525.3 K) but overestimated density, while SVR underestimated both. This work contributes a novel, interpretable MD–ML framework to integrate the inflection-point detection with data-driven model selection, offering a reproducible and accurate method for boiling point estimation that can be extended to other organic thermal systems.