Improving molecular force fields across configurational space by combining supervised and unsupervised machine...

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Machine Learning Force Fields (MLFFs) are crucial for molecular simulations.
• Current MLFFs often suffer from biased training data, limiting their accuracy to common configurations.
• This bias restricts the applicability of MLFF models in predictive molecular simulations.

Purpose of the Study:

• To develop a novel method for selecting training data to improve MLFF performance.
• To overcome the limitations of inhomogeneously distributed datasets in configurational space (CS).
• To enhance the accuracy and applicability of MLFFs for diverse molecular configurations, including non-equilibrium geometries.

Main Methods:

• Combined unsupervised and supervised machine learning (ML) methods.
• Clustering of configurational space (CS) into geometrically and energetically similar subregions.
• Iterative testing of MLFF performance on subregions and targeted inclusion of data from inaccurate regions into the training set.

Main Results:

• Achieved up to a twofold decrease in root mean squared errors for force predictions on non-equilibrium geometries.
• Demonstrated superior stability of ML models compared to default training approaches.
• Validated the approach across different ML methods, including kernel-based models (sGDML, GAP/SOAP) and deep neural networks (SchNet).

Conclusions:

• The proposed data selection strategy effectively mitigates bias towards common configurations.
• This method significantly widens the applicability range of MLFFs for molecular simulations.
• Enables reliable studies of processes involving highly out-of-equilibrium molecular configurations.