Challenges and opportunities for machine learning potentials in transition path sampling: alanine dipeptide and azobenzene studies

Nikita Fedik, Wei Li, Ying Wai Li, Benjamin Nebgen, Sergei TretiakDigital 2025 · DOI

Abstract

Machine Learning Interatomic Potentials (MLIPs) enable fast Transition Path Sampling (TPS) simulations, as demonstrated for alanine dipeptide. However, these models may fail in complex cases like azobenzene isomerization. This study highlights that adequate reference data and domain expertise are crucial for selecting test trajectories and ensuring the physical validity of the simulated pathways.

Introduction

In computational chemistry, studying "rare events"—such as protein conformational changes or molecular isomerization—presents a significant challenge. Standard Molecular Dynamics (MD) is often too slow to capture these transitions naturally. Transition Path Sampling (TPS) offers a solution by focusing on the transition region itself. However, TPS requires an efficient and accurate "engine" to calculate forces. While classical force fields are fast, they often lack the complexity to capture chemical transitions, and Quantum Mechanical (QM) methods are too computationally expensive for the thousands of trajectories required.

Conceptual overview of the integration of machine learning potentials with transition path sampling.

Machine Learning Interatomic Potentials (MLIPs) have emerged as a promising middle ground, acting as fast surrogates trained on QM data. This research investigates whether general-purpose MLIPs are ready for the rigors of TPS, specifically examining their performance in simple conformational changes versus complex bond-breaking events.

Methodology

The study evaluates two neural network-based MLIPs: HIP-NN-TS (utilizing "tensor sensitivity" for many-body information) and ANI-1x. These were benchmarked against classical force fields (Amber14, Sage) and Density Functional Theory (DFT) references. The core approach involved:

  • Transition Path Sampling (TPS): Employing a stochastic "shooting" approach to generate ensembles of reactive trajectories.
  • Potential Energy Surface (PES) Mapping: Comparing energy landscapes across different models to identify discrepancies in predicted pathways.
  • Manual Active Learning: Incorporating high-energy structures identified during TPS back into the training loop to evaluate model improvement and convergence.

Data and Benchmarking

The models were tested using several key datasets:

  • ANI-1x: A massive dataset of 5 million HCNO (Hydrogen, Carbon, Nitrogen, Oxygen) structures for initial training.
  • Alanine Dipeptide (AD): 10,000 snapshots from thermal MD for benchmarking, supplemented by 12,000 new configurations sampled from transition regions via TPS.
  • Azobenzene (AZ): Trajectories covering inversion (closed-shell) and rotation (open-shell) isomerization pathways.

Discussion: The Deception of RMSE and Model Physics

A critical insight from this work is that standard accuracy metrics, such as Root Mean Square Error (RMSE), can be deceptive. In the case of alanine dipeptide, ANI-1x exhibited a higher error than HIP-NN-TS but was more successful at identifying a specific minor energy basin ($a_L$) that HIP-NN-TS missed. This suggests that the underlying model architecture and the "physics" it captures are as important as the raw data.

Furthermore, the study identifies a "closed-shell trap." Most general MLIPs are trained on closed-shell DFT, which works for simple rotations but fails for molecules like azobenzene where the preferred transition path involves a biradical state. If the reference theory cannot describe the physics of bond-breaking, the ML model will inevitably fail to capture the correct pathway, regardless of the amount of training data.

Key Results

  • Performance Gains: MLIPs outperformed classical force fields by 300% in energy estimation and 900% in force estimation accuracy.
  • Chemical Accuracy: Both MLIPs achieved "chemical accuracy" (errors < 1 kcal/mol) for thermal MD of alanine dipeptide.
  • Azobenzene Limitations: For the rotational pathway of azobenzene, HIP-NN-TS predicted an unrealistically low barrier, while ANI-1x incorrectly treated the transition state as a local minimum. This failure was traced back to the limitations of the reference closed-shell DFT.
  • Active Learning Efficiency: Using TPS to identify transition structures for retraining proved highly effective. Adding just 0.2% more data from these regions reduced energy RMSE by approximately 30-35%.

Conclusion and Future Outlook

While MLIPs offer a powerful tool for accelerating rare-event simulations, they cannot be treated as black boxes. The "Garbage In, Garbage Out" principle applies: an MLIP cannot learn physics absent from its training data. Future work must focus on incorporating better reference theories (e.g., CASSCF or broken-symmetry DFT) to handle bond-breaking and integrating TPS into automated active learning loops to build reactivity-focused potentials. Domain expertise remains essential in choosing the right reference theory for the specific chemistry being studied.

Summary updated

Abstract

ML potentials enable fast TPS simulations ( e.g. , for alanine dipeptide) but may fail for complex cases like azobenzene isomerization. Adequate reference data and domain expertise are crucial for selecting test trajectories.
BibTeX
@inproceedings{fedik2025challenges,
  author = {Nikita Fedik and Wei Li and Ying Wai Li and Benjamin Nebgen and Sergei Tretiak and Ying Wai Li},
  title = {{Challenges and opportunities for machine learning potentials in transition path sampling: alanine dipeptide and azobenzene studies}},
  booktitle = {Digital},
  year = {2025},
  doi = {10.1039/d4dd00265b},
}
Figure