Accelerated Machine Learning Force Field for Predicting Thermal Conductivity of Organic Liquids

TL;DR

The paper tackles accurate, scalable prediction of thermal conductivity for organic liquids, a challenging transport-property problem. It introduces BAMBOO-TC, an MLFF workflow built on Graph Equivariant Differential Transformer (GEDT) with differential attention, trained on DFT data and calibrated by experimental densities. The method achieves a mean absolute percentage error of $14\%$ across 20 liquids, outperforming the conventional OPLS-AA force field, and shows transferability when densities are aligned with experiments. A Triton-based reimplementation yields up to $8\times$ inference speedups and $3.7\times$ overall MD speedups on modern GPUs, enabling rapid, large-scale predictions. Overall, the work demonstrates that integrating physics-grounded alignment with MLFFs can enable accurate, fast predictions of transport properties in molecular liquids and suggests a path toward improved ab initio-informed force fields.

Abstract

The thermal conductivity of organic liquids is a vital parameter influencing various industrial and environmental applications, including energy conversion, electronics cooling, and chemical processing. However, atomistic simulation of thermal conductivity of organic liquids has been hindered by the limited accuracy of classical force fields and the huge computational demand of ab initio methods. In this work, we present a machine learning force field (MLFF)-based molecular dynamics simulation workflow to predict the thermal conductivity of 20 organic liquids. Here, we introduce the concept of differential attention into the MLFF architecture for enhanced learning ability, and we use density of the liquids to align the MLFF with experiments. As a result, this workflow achieves a mean absolute percentage error of 14% for the thermal conductivity of various organic liquids, significantly lower than that of the current off-the-shelf classical force field (78%). Furthermore, the MLFF is rewritten using Triton language to maximize simulation speed, enabling rapid prediction of thermal conductivity.

Figures (6)

• Figure 1: Overview of the MLFF-based workflow for prediction of thermal conductivity. a: Schematic of the sequential training and prediction processes of the workflow. b: Schematic of the differential transformer inside the GEDT layers of the MLFF. c: Illustration of the reverse non-equilibrium molecular dynamics (rNEMD) for the prediction of thermal conductivity. The box is divided into slabs along z-axis, where a temperature gradient is created. The center slab is created as the hottest, both ends are created as the coolest. d: The temperature profile of PRA at 298K, including the upper and the lower half of the box in the simulation of tert-butanol.
• Figure 2: Prediction results of OPLS-AA and BAMBOO-TC before and after alignment. In both subfigures, predictions of OPLS-AA are shown in green circles, predictions of BAMBOO-TC before density alignment are shown in orange circles, predictions of BAMBOO-TC after density alignment are shown in blue circles. a: Correlation scatter comparing predicted densities to experimental densities. b: Correlation scatter comparing predicted thermal conductivities to experimental thermal conductivities. In all figures, the black line refers to $y=x$, and the standard deviation of the three independent replicates is used as the error bar.
• Figure 3: Assessment of transferability. a: Three categories of molecule: the BAMBOO-TC model presented in this work is initially trained with atomic clusters of aligned and trained molecules, and aligned with experimental data of aligned molecules. Zero-shot molecules do not participate in the training of BAMBOO-TC. b,d: Density (b) and thermal conductivity (d) predictions from BAMBOO-TC before density alignment. c,e: Density (c) and thermal conductivity (e) predictions from BAMBOO-TC after density alignment. In all figures, the black line refers to $y=x$, and the standard deviation of the three independent replicates is used as the error bar.
• Figure S1: Schematic of the a: GNN, b: GEDT layer inside the GNN, and c: differential transformer inside the GEDT layer.
• Figure S2: Fitting performance of BAMBOO-TC with GEDT and GET layers. GEDT layers fit the DFT PES better than GET layers, demonstrating higher R$^2$ scores and lower root mean square error (RMSE). a, b, c: BAMBOO-TC predicted energy, forces, and virial predictions with GEDT layers comparing with DFT labels in validation set. d, e, f: BAMBOO-TC predicted energy, forces, and virial predictions with GET layers comparing with DFT labels in validation set.