Synergistic integration of physical embedding and machine learning enabling precise and reliable force field - APNN

Aug 3, 2026

Upload by Ezra

Machine learning force fields can reproduce potential energy surfaces with quantum-chemical accuracy, yet still struggle with extrapolating to new chemical spaces, representing long-range electrostatics, and predicting challenging macroscopic properties. A physically informed neural network (PINN) is proposed by embedding physical constraints directly into neural-network parameters and applying global optimization. Using AMOEBA+ as the physics-based embedding model and the DEGDME dataset as a case study, the method achieves a precise, noise-robust ML force field with strong generalization and DFT-level accuracy. It enables efficient prediction of properties such as diffusion coefficient at minimal additional cost, offering a foundational PINN framework.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages21

Reading Time53 min

AccessFree to download

Introduction

Machine learning force fields and their limitations

Physics-informed approaches for ML force fields

Motivation and proposed method (APNN)

arXiv :2404 . 13368v1 [physics .chem-ph] 20 Apr 2024 Synergistic integration of physical embedding and machine learning enabling precise and reliable force field Lifeng Xu 1,2 and Jian Jiang 1,2* 1 Beijing National Laboratory for Molecular Sciences, State Key Laboratory of Polymer Physics and Chemistry, Institute of Chemistry, Chinese Academy of Sciences, Beijing, 100190, P.R., China. 2 University of Chinese Academy of Sciences, Beijing, 10049, P.R. , China. *Corresponding author(s). E-mail(s): [jiangj@iccas.ac.cn](jiangj@iccas.ac.cn) ; Abstract The machine learning force field has achieved significant strides in accurately reproducing the potential energy surface with quantum chemical accuracy. However, it still faces significant challenges, e.g., extrapolating to uncharted chemical spaces, interpreting long-range electrostatics, and mapping complex macroscopic properties. To address these issues, we advocate for a synergistic integration of physical principles and machine learning techniques within the framework of a physically informed neural network (PINN) . This innovative approach involves the incorporation of physical constraints directly into the parameters of the neural network, coupled with the implementation of a global optimization strategy. We choose the AMOEBA+ force field as the physics-based model for embedding, and then train and test it using the diethylene glycol dimethyl ether (DEGDME) dataset as a case study. The results reveal a significant breakthrough in constructing a precise and noise-robust machine learning force field. Utilizing two training sets with hundreds of samples, our model exhibits remarkable generalization and DFT accuracy in describing molecular interactions and enables a precise prediction of the macroscopic properties such as diffusion coefficient with minimal cost. This work provides a crucial insight into establishing a fundamental framework of PINN. 1 1 Introduction Molecular dynamics (MD) simulation is a pivotal tool for understanding the structure, property, and behavior of molecular systems, significantly advancing the frontiers of chemistry [1], biology [2, 3], and materials [4] science. Central to MD simulations is the force field (FF), which encodes interatomic interactions according to the chemical environment, thereby governing the motion and state of atoms. Attributed to the rapid and robust calculation of classical FF, MD enable exploration of systems comprising millions of atoms over microsecond timescales [5, 6], providing fundamental insight into thermodynamic and dynamic properties. Nevertheless, a FF based on simple physical expression usually imposes limitations on accurately capturing complex interactions such as charge penetration and polarization effects [7], resulting in notable deviations from reality. Hence, there has been a pressing demand for high-precision force fields to ensure the faithful rendition of the structural, thermodynamics, and dynamics features of real systems. In recent years, numerous machine learning (ML) force fields (FFs) achieving DFT accuracy in simulating interatomic potential have been proposed, e.g., DeepPotSE [8], DimeNet [9], PaiNN [10], GemNet-T [11], NequIP [12] and Allegro [13] MLFFs, etc [14] . The inherent training mechanism and flexible structure of neural network enable it to efficiently and accurately consider high-dimensional problems [15] such as many-body interaction [16] . However, in the absence of physical meaning, the traditional MLFFs are lack of extrapolative generalization, thus leading to significant model hallucination when applied to untrained data [14, 17–19] . Additionally, most of MLFFs only focus on learning local atomic environment within a cutoff distance [20] . The absence of long-range interaction renders MLFFs ineffective in simulating charged systems such as electrolytes [21], and proteins [22] . Fortunately, this challenge can be addressed by using a physically informed model b

FAQs

What main challenges do current machine learning force fields still face?

They often fail when extrapolating to uncharted chemical spaces, interpreting long-range electrostatics, and mapping complex macroscopic properties such as kinetic and dynamic quantities.

How does the proposed method combine physical principles with machine learning?

It uses a physically informed neural network that incorporates physical constraints directly into neural-network parameters and applies a global optimization strategy.

Why is AMOEBA+ used in this work, and what dataset is used for validation?

AMOEBA+ serves as the physics-based embedding model, and the DEGDME dataset is used as the case study to train and test the framework.

Aug 3, 2026

Upload by Ezra