Learning temperature‐dependent molecular representations for physicochemical property prediction

Abstract Physicochemical properties often display diverse and nonlinear dependencies on temperature. While machine learning has excelled at modeling molecular structural representations, accurately capturing the interplay between temperature and these properties remains a significant challenge. To address this, we introduce the Boltzmann layer, a neural network module inspired by the principles of statistical thermodynamics. Designed as a plug‐and‐play component, the Boltzmann layer enables seamless incorporation of temperature effects into molecular representations. By integrating the Boltzmann layer with the graph attention networks, our method achieves good predictive performance, yielding an average absolute relative deviation from 1.73% to 11.5% and an R 2 from 0.863 to 0.999 across 18 physicochemical properties. Notably, the model also excels in mixture systems, delivering exceptional performance on mixture viscosity and solubility. Our findings highlight that embedding physical domain knowledge within machine learning models leads to both greater predictive accuracy and deeper scientific insight.