Analytical gradients for molecular-orbital-based machine learning.

Area of Science:

• Computational chemistry
• Machine learning in quantum chemistry

Background:

• Molecular-orbital-based machine learning (MOB-ML) predicts energies but requires molecular orbitals.
• Accurate prediction of molecular properties necessitates efficient gradient calculations.

Purpose of the Study:

• To derive and implement MOB-ML analytical nuclear gradients within a general Lagrangian framework.
• To demonstrate the accuracy and efficiency of MOB-ML gradients for predicting molecular structures.

Main Methods:

• Formulation of MOB-ML gradients using a Lagrangian framework to satisfy orbital constraints.
• General applicability to various regression techniques and feature designs.
• Numerical validation on the ISO17 dataset.

Main Results:

• MOB-ML gradients achieve high accuracy comparable to traditional methods.
• Requires significantly less training data (energies only) compared to other ML approaches.
• Gradient evaluation cost is comparable to density-corrected DFT calculations.

Conclusions:

• MOB-ML analytical nuclear gradients offer a computationally efficient and accurate approach for quantum chemistry.
• This advancement enables accurate prediction of optimized molecular structures with reduced computational burden.