Efficient Long-Range Machine Learning Force Fields for Liquid and Materials Properties

John L. Weber, Rishabh D. Guha, Garvit Agarwal, Yujing Wei, Aidan A. Fike, Xiaowei Xie, James Stevenson, Biswajit Santra, Richard A. Friesner, Karl Leswing, Mathew D. Halls, Robert Abel, Leif D. Jacobson · 2025

Read Efficient Long-Range Machine Learning Force Fields for Liquid and Materials Properties by John L. Weber, Rishabh D. Guha, Garvit Agarwal, Yujing Wei, Aidan A. Fike, Xiaowei Xie, James Stevenson, Biswajit Santra, Richard A. Friesner, Karl Leswing, Mathew D. Halls, Robert Abel, Leif D. Jacobson (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-06462.txt.

Introduction

John Weber and a team of co-authors build fast force fields in twenty twenty-five. Machine learning force fields simulate atoms at near quantum accuracy for far less cost. Recent message passing models can span the whole periodic table. Long-range interactions and charge effects, though, are hard to include cheaply.

Their architecture, MPNICE, is an invariant message passing model. It predicts atomic partial charges step by step, which lets long-range forces enter naturally. Charge-dependent properties can then be predicted as well. Inference runs five to twenty times faster than models of comparable accuracy. They trained models for organic systems and benchmarked liquid and solid properties against experiment. New torsion scans with charged species add a fresh test set.

Catalog record

CREATORJohn L. Weber, Rishabh D. Guha, Garvit Agarwal, Yujing Wei, Aidan A. Fike, Xiaowei Xie, James Stevenson, Biswajit Santra, Richard A. Friesner, Karl Leswing, Mathew D. Halls, Robert Abel, Leif D. Jacobson
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE101 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-06462

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Efficient Long-Range Machine Learning Force Fields for Liquid and Materials Properties John L. Weber, Rishabh D. Guha, Garvit Agarwal, Yujing Wei, Aidan A. Fike, Xiaowei Xie, James Stevenson, Biswajit Santra, Richard A. Friesner, Karl Leswing, Mathew D. Halls, Robert Abel, Leif D. Jacobson Abstract Machine learning force fields (MLFFs) have emerged as a sophisticated tool for cost-efficient atomistic simulations approaching DFT accuracy, with recent message passing MLFFs able to cover the entire periodic table. We present an invariant message passing MLFF architecture (MPNICE) which iteratively predicts atomic partial charges, including long-range interactions, enabling the prediction of charge-dependent properties while achieving 5-20x faster inference versus models with comparable accuracy. We train direct and delta-learned MPNICE models for organic systems, and benchmark against experimental properties of liquid and solid systems. We also benchmark the energetics of finite systems, contributing a new set of torsion scans with charged species and a new set of DLPNO-CCSD(T) references for the TorsionNet500 benchmark. We additionally train and benchmark MPNICE models for bulk inorganic crystals, focusing on structural ranking and mechanical properties. Finally, we explore multi-task models for both inorganic and organic systems, which exhibit slightly decreased performance on domain-specific tasks but surprising generalization, stably predicting the gas phase structure of ≃ 500 \simeq 500 Pt/Ir organometallic complexes despite never training to organometallic complexes of any kind. keywords American Chemical Society, L a T e X † † affiliation: Schrödinger Inc., 1540 Broadway, 24 th floor, New York, NY 10036 † † affiliation: Schrödinger Inc., 1540 Broadway, 24 th floor, New York, NY 10036 † † affiliation: Schrödinger Inc., 1540 Broadway, 24 th floor, New York, NY 10036 † † affiliation: Schrödinger Inc., 1540 Broadway, 24 th floor, New York, NY 10036 † † affiliation: Department of Chemistry, Columbia University, 3000 Broadway, New York, NY 10027 † † affiliation: Schrödinger Inc., 1540 Broadway, 24 th floor, New York, NY 10036 † † affiliation: Schrödinger Inc., 101 SW Main Street, Suite 1300, Portland, OR 97204 † † affiliation: Schrödinger Inc., 1540 Broadway, 24 th floor, New York, NY 10036 † † affiliation: Schrödinger Inc., 1540 Broadway, 24 th floor, New York, NY 10036 † † affiliation: Department of Chemistry, Columbia University, 3000 Broadway, New York, NY 10027 † † affiliation: Schrödinger Inc., 1540 Broadway, 24 th floor, New York, NY 10036 † † affiliation: Schrödinger Inc., 9868 Scranton Road, Suite 3200 San Diego, CA 92121 † † affiliation: Schrödinger Inc., 1540 Broadway, 24 th floor, New York, NY 10036 † † affiliation: Schrödinger Inc., 101 SW Main Street, Suite 1300, Portland, OR 97204 † † email: [email protected] † † abbreviations: MPNICE † † suppinfo: 1 Introduction Density Functional Theory (DFT) is the current workhorse of computational modeling of condensed phase systems. Parametrized functionals have achieved near chemical accuracy for select applications, Mardirossian and Head-Gordon (2017) while retaining high enough efficiency such that ab initio molecular dynamics (AIMD) simulations using DFT can reliably be run on tens to hundreds of atoms. For properties which require larger simulations or more than a few picoseconds to resolve, empirically parametrized interatomic potentials, also known as classical force fields, have been used for decades, Jorgensen et al. (1996) ; Monticelli and Tieleman (2013) and are efficient enough to run hundreds of nanoseconds of molecular dynamics (MD) a day, at the expense of reduced accuracy and generalization. While parameterizing classical force fields for specific application domains, such as organic druglike molecules, has had its fair share of success, there are many notable areas where classical force fields are lacking, such as treating reactive systems and diverse inorganic materials. In recent decades, machine learning force fields (MLFFs), sometimes referred to as machine learning interatomic potentials (MLIPs), have increased the capacity of force fields by abstracting the functional form and dramatically increasing the number of learnable parameters. MLFFs have been shown to effectively reproduce DFT accuracy for targeted applications at a fraction of the cost of DFT, albeit on the order of 100 times more expensive than classical force fields. High dimensional neural network potentials (HDNNPs) were the first MLFFs demonstrated to achieve high accuracy on large organic datasets Smith et al. (2017) ; Smith et al. (2018) ; Stevenson et al. (2019) ; Smith et al. (2020) . Such architectures exhibit poor scaling with the number of supported elements due to the use of element-dependent Behler-Parinello type symmetry functions. Behler and Parrinello (2007) As such, efforts to produce “general” models of this type were restricted to subsets of the periodic table, typically neutral organic molecules Smith et al. (2020) . In order to distinguish different charge states, as well as include long-range electrostatics interactions, some models additionally predict fractional charges on each atom. Ko et al. (2021) ; Jacobson et al. (2022) ; Anstine et al. (2025) Models such as these have been used to perform high-fidelity simulations of specific organic systems, including redox chemistry and electrolytes for Li-ion batteries. Dajnowicz et al. (2022) ; Kocer et al. (2024) The development of message passing networks Schütt et al. (2017) ; Unke and Meuwly (2019) has effectively solved the scaling of MLFFs with respect to the number of chemical elements by embedding the element in an atomic feature vector. There has also been rapid growth in the number and scale of public databases for materials properties calculated using DFT, including the Materials Project Jain et al. (2013) and the Open Catalyst Project. Chanussot et al. (2021) ; Tran et al. (2023) ; Barroso-Luque et al. (2024) The combination of flexible models and diverse datasets has enabled the construction of general-purpose MLFFs for inorganic materials, with a wide variety of architectures and models showcasing impressively high accuracy and generalization to unseen structures. Deng et al. (2023) ; Chen and Ong (2022) ; Batatia et al. (2023) ; Fu et al. (2025) ; Barroso-Luque et al. (2024) ; Park et al. (2024) ; Choudhary and DeCost (2021) ; Bochkarev et al. (2024) ; Merchant et al. (2023) ; Neumann et al. (2024) ; Zhang et al. (2024) With the introduction of public benchmarks and leaderboards, Riebesell et al. (2023) model development has trended towards larger models, often incorporating equivariant geometric features in order to increase the data-efficiency of training at the expense of increased inference cost. Loew et al. (2024) While development of general models in the inorganic materials landscape has advanced significantly in the past two years, it has yet to make significant contact with general organic datasets, perhaps due to the significant differences in the levels of theory employed. The majority of inorganic datasets are trained on DFT data generated with the PBE functional, utilizing pseudopotentials and plane wave basis sets under periodic boundary conditions, whereas for organic molecules, all-electron calculations are used, often employing hybrid or long-range corrected functionals combined with an atom-centered basis set. The reasons behind these choices are fundamental, involving the inaccuracy of pure functionals for organic potential energy surfaces (PES) Mardirossian and Head-Gordon (2017) ; Dreuw and Head-Gordon (2005) and the problem of Hartree-Fock exchange in hybrid functionals introducing errors when treating zero-gap systems. Paier et al. (2007) As a result, however, the absolute energy scale of the training data is different between organic and inorganic datasets, leading to significant conflicts

Read the complete text