Machine learning for potential energy surfaces: An extensive database and assessment of methods

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Accurate potential energy surfaces (PES) are crucial for understanding molecular behavior and reaction dynamics.
• Traditional methods for PES construction can be computationally expensive and time-consuming.
• Machine learning (ML) offers a promising avenue for accelerating PES development.

Purpose of the Study:

• To assess the performance of various ML algorithms for predicting molecular electronic energies.
• To evaluate the black-box nature, robustness, and efficiency of different ML models in PES construction.
• To explore delta-learning strategies for enhancing prediction accuracy.

Main Methods:

• Construction of an extensive database with ~10.5 million configurations for 15 small molecules.
• Computation of electronic energies using SCF, RI-MP2, RI-MP2-F12, and CCSD(F12*)(T) methods.
• Implementation and evaluation of ML algorithms including Gaussian Process Regression (GPR), Kernel Ridge Regression, Support Vector Regression, and Neural Networks (NNs).
• Exploration of GPR variants (sparse, Markov Chains) and NN architectures.
• Application of delta-learning strategies.

Main Results:

• Various ML algorithms demonstrate capability in predicting molecular energies for PES construction.
• Delta-learning strategies, especially combining RI-MP2 with ML-predicted CCSD(F12*)(T)-RI-MP2 differences, show significant promise.
• The study provides insights into the trade-offs between different ML methods regarding accuracy, robustness, and efficiency.

Conclusions:

• ML algorithms are viable tools for constructing accurate potential energy surfaces.
• Delta-learning approaches enhance the efficiency and accuracy of ML-based energy predictions.
• The findings pave the way for more automated and efficient PES construction for anharmonic vibrational computations.