Chinese Scientists Resolve Structural Collapse in Organic Simulations Using Physics-Embedded Machine Learning Models

Chinese Academy of Sciences researchers improve organic molecular simulation accuracy by 88% using new physics-guided machine learning force fields.

By: AXL Media

Published: Apr 28, 2026, 9:25 AM EDT

Source: Information for this report was sourced from EurekAlert!

Overcoming Data Limitations in Organic Molecular Dynamics

The application of machine learning force fields (MLFF) in organic chemistry has long been hindered by structural instability and inaccurate macroscopic predictions. Traditional data-driven models often struggle with the complex interplay of intramolecular covalent bonds and intermolecular van der Waals forces, leading to "structural collapse" where bonds break or atoms collide during simulations. To address this, Professor Jian Jiang and his team at the Institute of Chemistry, Chinese Academy of Sciences, have introduced a hybrid approach that embeds physical laws directly into the machine learning framework. This methodology shifts the reliance away from massive datasets, focusing instead on the fundamental physical equations that govern molecular behavior.

Adaptive Sampling to Prevent Molecular Structural Failure

The first major advancement presented by the team is a physics-guided adaptive bond length sampling method. By reading topology files from empirical force fields, the model can precisely distinguish between different atom and bond types within complex organic environments. This allows the system to determine the ideal range and probability for sampling chemical bonds, specifically targeting high-energy regions that purely data-driven models typically ignore. In tests involving fluorinated engineering fluids and acetaminophen, this adaptive approach eliminated structural collapse probabilities that were as high as 77% in standard models, ensuring stability across 100 consecutive high-temperature simulation tests.

Bridging the Gap Between Microscopic and Macroscopic Data

Even when machine learning models accurately capture microscopic atomic forces, they frequently fail to predict macroscopic properties like density and viscosity. The researchers addressed this discrepancy by embedding a top-down correction strategy based on the DFT-CSO dispersion equation. This physical equation allows for targeted adjustments to intermolecular interactions using a single damping parameter optimized against experimental density data. By grounding the model in observable macroscopic reality, the team has created a system that corrects for the inherent systematic biases found in standard quantum chemistry reference methods.