Challenges for machine learning force fields in reproducing potential energy surfaces of flexible molecules

Area of Science:

• Computational Chemistry
• Machine Learning in Molecular Dynamics
• Physical Chemistry

Background:

• Flexible molecule dynamics arise from local bond fluctuations and long-range interactions, creating complex potential energy surfaces (PESs).
• Accurately modeling these PESs is crucial for understanding molecular behavior but challenging due to their complexity.

Purpose of the Study:

• To evaluate the performance of state-of-the-art machine learning (ML) models in reproducing complex PESs using limited reference data.
• To benchmark ML models using the cis-trans thermal relaxation of azobenzene as a case study, considering multiple transition mechanisms.

Main Methods:

• Assessed sGDML, SchNet, Gaussian Approximation Potentials/Smooth Overlap of Atomic Positions (GAPs/SOAPs), and Behler-Parrinello neural networks.
• Utilized a limited dataset for training and evaluated model performance on reproducing the potential energy surface (PES).
• Employed the cis-to-trans thermal relaxation of azobenzene as a benchmark system involving at least three transition mechanisms.

Main Results:

• Models like GAPs/SOAPs, SchNet, and sGDML achieved chemical accuracy (1 kcal mol⁻¹) with fewer than 1000 data points.
• Model predictions varied significantly based on the ML method and the specific region of the PES sampled.
• Discrepancies were observed in predictions between equilibrium and transition regions, and across different transition mechanisms.

Conclusions:

• Current ML models face challenges in accurately representing complex PESs, largely due to limitations in atom-based descriptors.
• A single ML model may not be sufficient for learning the entire PES; specialized local models are recommended.
• Future approaches should consider using multiple local models with tailored descriptors, training sets, and architectures for different PES regions.