Bridging Quantum Mechanics to Organic Liquid Properties via a Universal Force Field

TL;DR

This work presents ByteFF-Pol, a graph neural network–parameterized polarizable force field trained exclusively on high-level QM data to predict macroscopic liquid properties. By enforcing a physically motivated energy decomposition and aligning force-field terms with ALMO-EDA labels, ByteFF-Pol achieves zero-shot transferability across diverse organic liquids and electrolytes, outperforming traditional and ML force fields in density, evaporation enthalpy, viscosity, and conductivity benchmarks. The method enables efficient MD simulations (≈40 ns/day on 10{,}000-atom systems) while maintaining ab initio–level fidelity, bridging microscopic QM calculations and macroscopic properties essential for electrolyte design and solvent optimization. This framework advances data-driven materials discovery by enabling rapid exploration of large chemical spaces with transferable, physically grounded force fields.

Abstract

Molecular dynamics (MD) simulations are essential tools for unraveling atomistic insights into the structure and dynamics of condensed-phase systems. However, the universal and accurate prediction of macroscopic properties from ab initio calculations remains a significant challenge, often hindered by the trade-off between computational cost and simulation accuracy. Here, we present ByteFF-Pol, a graph neural network (GNN)-parameterized polarizable force field, trained exclusively on high-level quantum mechanics (QM) data. Leveraging physically-motivated force field forms and training strategies, ByteFF-Pol exhibits exceptional performance in predicting thermodynamic and transport properties for a wide range of small-molecule liquids and electrolytes, outperforming state-of-the-art (SOTA) classical and machine learning force fields. The zero-shot prediction capability of ByteFF-Pol bridges the gap between microscopic QM calculations and macroscopic liquid properties, enabling the exploration of previously intractable chemical spaces. This advancement holds transformative potential for applications such as electrolyte design and custom-tailored solvent, representing a pivotal step toward data-driven materials discovery.

Figures (4)

• Figure 1: Framework of ByteFF-Pol Given molecular graphs, the ByteFF GNN model predicts both bonded and non-bonded force field parameters. During the training stage, these parameters are combined with atomic coordinates in energy functions to compute decomposed energy terms, which are then fitted to reference labels from the ALMO-EDA calculations. Once trained, the force field parameters are obtained in a single inference step, and are used for simulations in standard MD softwares.
• Figure 2: ByteFF-Pol is validated against DFT references. (a) Comparison of interaction energies predicted by ByteFF-Pol versus DFT on the validation set (a random sampled subset of 76 k conformations). (b, c) Dimer scan of (b) two EMC molecules and (c) an EMC molecule with a lithium ion (solid lines for ByteFF-Pol and dashed lines for DFT). (d) Cluster of a lithium ion surrounded by three DMC molecules and a FSI$^-$ anion. The conformations were first relaxed by DFT (light blue carbon) and then by ByteFF-Pol (grey carbon). (e, f) Decomposed binding energies of clusters with neutral molecules only (e) and clusters with neutral molecules and ions (f). The energy terms are accumulated so that the final point equals to the overall binding energy (solid lines for ByteFF-Pol and dashed lines for DFT). The corresponding components and conformations of the clusters are provided in Supplementary Section 2.1. The full names and SMILES strings corresponding to the molecular abbreviations are listed in Supplementary Table 1.
• Figure 3: Benchmarks on the thermodynamic properties of molecular liquids. Comparison of different force fields on (a) density and (b) evaporation enthalpy of pure molecular liquids from QUBEKit hortonQUBEKitAutomatingDerivation2019; (c) density of binary molecular liquids from ThermoML chirico2003thermoml; and (d) density prediction by ByteFF-Pol on additional pure molecular liquids from CRC handbook Lide2004CRC. The results of MACE-OFF were acquired from Ref kovacsMaceOff2025. Pearson correlation coefficients (r) of each force field are provided in the legends.
• Figure 4: Evaluation of ByteFF-Pol for predicting electrolyte properties. (a-c) Comparison of different force fields in predicting (a) density, (b) viscosity and (c) conductivity of liquid electrolytes, respectively. Pearson correlation coefficients (r) of each force field are provided in the legends. (d-e) Examples for conductivity at (d) varying lithium salt concentrations; (e) at different temperatures (solid lines denote simulation results, dashed lines denote experimental data). The numbers in the legends indicate the approximate molar fraction of each species. (f) Examples for self-diffusion coefficient ($D$) of different species in electrolyte. Upper bars correspond to $D$ of species X (labeled on the x axis) in the X:EMC binary solvent (molar ratio=1:1); and lower bars show the $D$ of X in the LiPF6:X:EMC electrolytes (molar ratio=1:4:4). Experimental data are obtained from C9EE00141G. The full names and SMILES strings corresponding to the molecular abbreviations are provided in Supplementary Table 1.