Quasiclassical Trajectory Simulation as a Protocol to Build Locally Accurate Machine Learning Potentials.

Area of Science:

• Computational Chemistry
• Chemical Dynamics
• Machine Learning

Background:

• Direct trajectory calculations are computationally expensive for mechanistic studies.
• Machine learning-based potential energy surfaces (ML-PES) offer a computationally efficient alternative.
• Developing robust ML-PES requires extensive training data covering broad configuration spaces.

Purpose of the Study:

• To demonstrate an efficient strategy for constructing locally accurate ML-PES.
• To leverage quasiclassical trajectory (QCT) calculations for localized sampling.
• To reduce the computational burden of ab initio trajectory calculations in mechanistic explorations.

Main Methods:

• Utilized quasiclassical trajectory (QCT) calculations for localized sampling of configuration space.
• Developed machine learning-based potential energy surfaces (ML-PES) using the sampled data.
• Applied the method to two model reactions: methyl migration of i-pentane cation and cyclopentadiene dimerization.

Main Results:

• Locally accurate ML-PESs were efficiently obtained.
• The developed ML-PESs accurately reproduced static and dynamic features of the model reactions.
• Key properties like time-resolved free energy and entropy changes were successfully reproduced.

Conclusions:

• Localized sampling with QCT enables efficient construction of accurate ML-PES.
• This approach significantly reduces computational cost while maintaining accuracy for specific reaction properties.
• The method is robust for reproducing static and dynamic aspects of chemical reactions.