[論文レビュー] Efficient Long-Range Machine Learning Force Fields for Liquid and Materials Properties

John L. Weber, Rishabh D. Guha+11 more|ArXiv.org|May 9, 2025

Machine Learning in Materials Science被引用数 4

ひとこと要約

This paper introduces MPNICE, a charge-aware invariant message-passing ML force field that incorporates long-range electrostatics via iterative charge equilibration, enabling accurate predictions for organic and inorganic systems with much faster inference and strong generalization.

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 $\simeq500$ Pt/Ir organometallic complexes despite never training to organometallic complexes of any kind.

研究の動機と目的

• Develop a charge-aware invariant message passing ML force field (MPNICE) that can model long-range interactions in diverse materials.
• Enable fast, scalable inference (5–20x faster) while maintaining DFT-like accuracy across organic and inorganic systems.
• Benchmark MPNICE on organic molecules, ions, inorganic crystals, and hybrid datasets against relevant references and existing models.
• Demonstrate capabilities in predicting conformational energies, tautomer energies, crystal structure ranking, and basic liquid properties, including MD compatibility.

提案手法

• Use an invariant message passing architecture with iterative interaction blocks to propagate information among atoms within a cutoff.
• Predict atomic partial charges at each interaction step via an approximation to charge equilibration (Qeq) and use these charges in subsequent message passing.
• Represent local environments with atomic environment vectors (AEVs) including a charge-weighted radial term (QRAEV) to encode long-range Coulombic effects.
• Decompose total energy into atom-wise contributions plus explicit electrostatics and optional dispersion (D3) terms; share hidden representations across multiple theory heads for multitask learning.
• Train direct and delta-learned heads for organic and inorganic data, including multi-task setups that combine different reference levels of theory.
• Provide a final energy readout per atom after T interaction blocks and allow zero-shot predictions and cross-domain transfer.

Figure 1 : (a) An overview of the MPNICE architecture. Atomic features are initialized with an embedding of the chemical element, whereas geometric environments are represented using a set of atomic environment vectors (AEVs). Atomic features, atomic charges (initialized to zero), and AEVs are used

実験結果

リサーチクエスチョン

• RQ1 Can a charge-aware invariant MP framework with iterative charge equilibration accurately capture long-range electrostatics across organic and inorganic systems?
• RQ2 What is the performance of MPNICE on conformational energetics, tautomer energies, crystal structure ranking, and MD-relevant properties compared to existing transferable MLFFs?
• RQ3 How well does MPNICE generalize in zero-shot scenarios, including organometallic-like complexes, and under multi-task training across disparate datasets?
• RQ4 How do different output heads (levels of theory) affect accuracy in organic vs inorganic benchmarks?
• RQ5 What are the inference speed benefits of MPNICE relative to comparable models at similar accuracy?

主な発見

• MPNICE achieves accurate, charge-aware predictions with 5–20x faster inference than models with comparable accuracy.
• Delta-learned and multi-task MPNICE variants reach high accuracy on torsion scans, often outperforming prior models on relevant datasets.
• Organic_MPNICE_TB and Organic_MPNICE substantially improve tautomer energy prediction, reaching near-chemical accuracy in some tests.
• Organic_Crystals_MPNICE with appropriate heads ranks organic crystal structures with improved lattice energy and density predictions versus several baselines.
• Zero-shot capabilities extend to organometallic-like Pt/Ir complexes, indicating robust generalization beyond training domains.
• Molecular dynamics and bulk property benchmarks (e.g., densities, diffusion in amorphous LiAlO2, and NaCl phonon corrections) show strong agreement with experiments and high-level references.
• The framework supports long-range electrostatics through Qeq-based charge equilibration, enabling correct charge-dependent properties across datasets.
• Hybrid models demonstrate stable performance by sharing representations while using dataset-specific energy heads.

Figure 2 : Violin plots showing the distribution of relative energy RMSDs for separate torsion scans in the Torsion2000 test set are given in panel (d), the line shows the median RMSD error. Panels (a)-(c) highlight three examples from models that give overall good performance on torsion scans. (a)