Artificial Intelligence in Chemistry: Developing Better Machine Learning Models for Potential Energy Surfaces
Potential energy surfaces (PESs) are important in understanding molecular behavior, chemical reactions, and material properties. They show how potential energy of a system changes as the positions of its atoms or molecules change. However, they can be complex and challenging to accurately compute, especially for large molecules or systems. The reliability of machine learning (ML) models depends heavily on the diversity of training data, particularly for chemically reactive systems.
Formulating a balanced and diverse dataset for a reactive system is difficult and can lead to ML models with overfitting problems. Researchers at the University of California, Lawrence Berkeley National Laboratory, and Penn State University developed an active learning workflow that expands an originally formulated Hydrogen combustion dataset by preparing collective variables (CVs) for systematic sampling.
Using this active learning strategy, the researchers were able to create a more diverse and balanced ML model for hydrogen combustion. The ML models were able to accurately predict the change in the transition state and reaction mechanism at finite temperature and pressure for hydrogen combustion. The researchers also used metadynamics as an efficient sampling tool for unstable structures, which helped with the active learning workflow.
The use of metadynamics allowed the researchers to identify holes in the PES landscape and inform the ML model through retraining with this data. This approach can help solve the problem of potentially unreliable predictions that ML models might face. The researchers used the active learning approach on an example where the ML model performance was tracked, and they are considering alternate approaches for future work.
Check out the published paper to learn more about this research. Don’t forget to also join the ML SubReddit, Facebook Community, Discord Channel, and Email Newsletter to stay updated on the latest AI news and projects.
If you like this work, you will love our newsletter.