Quasi-Classical Trajectory Calculation of Rate Constants Using an Ab Initio Trained Machine Learning Model (aML-MD)...

Area of Science:

• Computational Chemistry
• Materials Science
• Chemical Physics

Background:

• Machine learning (ML) models offer improved accuracy in molecular dynamics (MD) simulations.
• Training ML models requires large, accurate datasets, which are computationally expensive to generate.
• Inconsistent data can lead to unreliable ML models that fail to capture underlying physics.

Purpose of the Study:

• To develop ab initio trained ML-based MD (aML-MD) models using transfer learning.
• To leverage multifidelity data from Density Functional Theory (DFT) and multireference calculations.
• To improve the accuracy, efficiency, and generalization of ML models in MD.

Main Methods:

• Utilized transfer learning within the Deep Potential MD framework.
• Trained aML-MD models using DFT and multireference data of varying accuracy.
• Calculated rate constants for the H + HO2 reaction using quasi-classical trajectories.

Main Results:

• The aML-MD model with transfer learning accurately predicted rate constants for the H + HO2 reaction.
• Achieved a computational cost reduction of over five times compared to using only high-accuracy quantum chemistry data.
• Demonstrated the effectiveness of multifidelity data in improving ML model performance.

Conclusions:

• Transfer learning enables the development of accurate and efficient aML-MD models.
• Multifidelity data significantly reduces the computational cost of generating training sets for ML potentials.
• This approach holds great potential for advancing molecular dynamics simulations.