Acceleration of saddle-point searches with machine learning

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning Applications

Background:

• Saddle point identification on potential-energy surfaces (PES) is crucial for understanding molecular transitions.
• Traditional methods rely heavily on numerous computationally expensive ab initio calculations.
• Current approaches often waste resources on non-critical electronic structure calculations.

Purpose of the Study:

• To introduce a machine learning (ML) approach for accelerating saddle point searches.
• To reduce the number of computationally intensive ab initio force calls required.
• To enhance the efficiency of exploring potential energy surfaces.

Main Methods:

• Developing machine learning models trained on atomistic simulation data.
• Conducting saddle point searches within the ML representation for speed.
• Using ab initio calculations to verify ML predictions and identify areas for model improvement.

Main Results:

• Demonstrated a dramatic reduction in the number of ab initio force calls needed.
• Successfully applied the ML approach to two example problems.
• Showcased the iterative improvement of ML models with new training data.

Conclusions:

• ML offers a powerful strategy to accelerate saddle point searches in atomistic simulations.
• This method significantly reduces computational overhead compared to traditional techniques.
• Wider adoption of ML is expected to revolutionize exploration of potential energy surfaces.