Hierarchical machine learning of potential energy surfaces

Area of Science:

• Computational Chemistry
• Materials Science
• Quantum Chemistry

Background:

• Accurate potential energy surfaces (PESs) are crucial for understanding molecular behavior.
• Traditional methods for generating PESs are computationally expensive.
• Machine learning offers a promising avenue for accelerating PES calculations.

Purpose of the Study:

• To present a novel hierarchical machine learning (hML) approach for constructing highly accurate PESs.
• To optimize the training set size and composition for constituent machine learning models.
• To minimize computational cost while achieving desired accuracy.

Main Methods:

• Utilizing multiple delta-machine learning models trained on energies and energy corrections from a hierarchy of quantum chemical methods.
• Employing a semi-automatic procedure for optimal training set determination.
• Building machine learning models with kernel ridge regression and structure-based sampling for training points.

Main Results:

• Demonstrated significant reduction in computational cost (factor of 100) for generating a CH3Cl PES.
• Achieved high accuracy with errors of approximately 1 cm⁻¹.
• Validated the effectiveness of the hML approach for ab initio PES construction.

Conclusions:

• The hML method provides a computationally efficient pathway to accurate potential energy surfaces.
• This approach enables faster and more cost-effective simulations in computational chemistry.
• Hierarchical machine learning represents a significant advancement in PES generation techniques.