A groundbreaking advancement in the field of molecular dynamics simulations has been unveiled by Jian Jiang and his team at the Institute of Chemistry, Chinese Academy of Sciences. Their pioneering research addresses longstanding challenges associated with machine learning force fields (MLFFs) when applied to organic molecular systems. Specifically, the scientists tackled critical issues like molecular structure collapse during extended simulations and inaccuracies in predicting macroscopic properties such as density and viscosity. By integrating physical principles directly into the development and refinement of MLFFs, they devised two innovative physical embedding techniques that dramatically enhance simulation stability and predictive accuracy, even with limited training data.
Molecular dynamics simulations serve as indispensable tools to explore the behavior of chemical, biological, and material systems at atomistic scales. Traditional approaches include high-precision ab initio molecular dynamics, which, although accurate, demand vast computational resources and remain unsuitable for large or long-duration simulations. Conversely, classical empirical force fields provide computational efficiency but often compromise accuracy, especially for complex organic molecules featuring diverse interaction types. MLFFs have emerged as a promising compromise, leveraging data-driven models to replicate the potential energy surfaces with greater accuracy while maintaining efficiency. Nevertheless, MLFFs struggle to concurrently and accurately represent both intramolecular covalent bonds and weaker intermolecular van der Waals forces, producing unstable simulations or erroneous macroscopic predictions.
The core difficulty lies in the dual nature of organic systems: intramolecular interactions dominated by strong covalent bonds require capture of high-energy states, while intermolecular forces must be characterized with sufficient fidelity to replicate bulk properties. Conventional MLFF training, driven predominantly by quantum mechanical reference data, often neglects extreme bond distortions, causing structural failures like unphysical bond breaking during prolonged simulations. Meanwhile, although the model’s microscopic predictions—energies, forces, radial distribution functions—may appear accurate, macroscopic thermodynamic properties can sharply diverge from experimental values due to insufficient modeling of intermolecular interactions.
To surmount these obstacles, Jiang’s team proposed a two-step physical embedding framework. The first step is a physics-guided adaptive bond length sampling method. This technique integrates empirical force field knowledge, specifically topology files outlining atom and bond classifications along with bond force constants, into the data sampling pipeline. Unlike traditional uniform bond stretching approaches, the adaptive method strategically targets high-energy bond regions that are typically underrepresented yet prone to structural collapse, assigning sampling probabilities based on physical bond stiffness. This nuanced sampling ensures comprehensive coverage of critical configurational spaces without inducing unphysical artifacts such as self-consistent field (SCF) non-convergence or anomalous forces.
Empirical verification was performed on three representative organic systems: fluorinated engineering fluids, alanine tripeptides, and acetaminophen molecules. With only 50 single-molecule samples per system used for model training and validation, the original MACE MLFF model exhibited collapse probabilities of 59%, 22%, and 77%, respectively. Following integration of the adaptive bond length sampling, the refined MLFFs passed 100 independent 100-picosecond high-temperature molecular dynamics runs without structural failure. These results illustrate a significant leap in the long-term stability and robustness of MLFFs, accomplished with minimal data augmentation.
The second innovation addresses the persistent challenge of accurately predicting macroscopic properties tied to intermolecular forces. Jiang’s group introduced a top-down correction mechanism rooted in physical equation embedding. This approach leverages the DFT-Corrected Screening Overlap (DFT-CSO) dispersion equation, embedding it into the MLFF architecture as an adjustable correction term. By tuning a damping parameter to modulate the strength of dispersion interactions, the method aligns intermolecular potentials to reproduce experimental densities, effectively compensating not just for MLFF fitting errors but also systematic biases inherent in the underlying quantum mechanical references.
Applied to battery electrolyte solvents—specifically mixtures of ethylene carbonate (EC) and methyl ethyl carbonate (EMC) as well as pure EMC systems—this correction yielded substantial improvements in predictive accuracy. Density prediction errors dropped by 78% and 88% for two model variants, reaching deviations as low as 0.006 and 0.012 g/cm³ respectively. Viscosity accuracy similarly improved by 38% and 77%, highlighting the technique’s capacity to enhance kinetic property predictions at negligible additional computational cost. Parameter optimization required only hours, underscoring the approach’s practicality for rapid model refinement and deployment.
Interpretability studies reinforced the physical significance of the embedded correction. Volume scan analyses showed shifts in interaction potential minima post-correction, validating the method’s ability to enhance or attenuate intermolecular forces directionally. Notably, atomic force root mean square errors changed by less than 0.8 meV/Å, well within the model’s intrinsic fitting uncertainty, confirming that macroscopic improvements stemmed from subtle force adjustments. The radial distribution functions remained essentially unchanged, indicating that intramolecular structural representations were preserved and that corrections selectively improved intermolecular interaction fidelity.
This dual-strategy framework—comprising physics-aware adaptive data sampling and physically grounded model post-processing—provides a scalable, low-cost solution to longstanding bottlenecks in MLFF construction. It bridges the gap between data-driven modeling and fundamental physical chemistry, yielding MLFFs that are simultaneously precise, interpretable, and transferable across diverse organic molecular systems. Beyond enhancing simulation accuracy, the methods enable high-fidelity studies of engineering fluids, peptides, pharmaceutical compounds, and organic solvents under practical computational budgets.
Breaking with conventional approaches that typically increase dataset size or model complexity to achieve improved performance, this work highlights the profound benefits of embedding domain knowledge directly into machine learning pipelines. Such embedding not only reduces the data dependency but also supports rapid model adaptation and fine-tuning for new chemical environments. Prospects for future work include expanding the correction framework with additional tunable physical parameters and extending its reach to efficiently capture kinetic phenomena such as viscosity, promising further gains in predictive power and applicability.
This seminal research was published as an open-access article in CCS Chemistry, the flagship journal of the Chinese Chemical Society. The undertaking was supported by the National Natural Science Foundation of China and the Strategic Priority Research Program of the Chinese Academy of Sciences. By marrying advanced computational modeling with deep physical insight, Jian Jiang and colleagues have charted a promising path toward more reliable and generalizable molecular simulations, with broad implications for chemistry, materials science, and drug discovery.
Subject of Research: Not applicable
Article Title: Physical Embedding Machine Learning Force Fields for Organic Systems
News Publication Date: 6-Mar-2026
Web References:
https://www.chinesechemsoc.org/journal/ccschem
http://dx.doi.org/10.31635/ccschem.026.202506780
References: Jiang J. et al. CCS Chem., 2025, 7(3): 716-730.
Image Credits: CCS Chemistry
Keywords
Machine learning, force fields, molecular dynamics, physical embedding, organic systems, adaptive bond length sampling, intermolecular interactions, DFT-CSO correction, macroscopic property prediction, simulation stability, quantum chemistry, computational modeling
