Multifidelity Neural Network Formulations for Prediction of Reactive Molecular Potential Energy Surfaces.

Area of Science:

• Computational Chemistry
• Machine Learning in Science
• Quantum Mechanics

Background:

• Accurate prediction of molecular properties requires high-fidelity quantum chemistry simulations, which are computationally expensive.
• Existing methods often rely on single levels of theory, limiting efficiency and accuracy when computational resources are constrained.
• Neural networks offer a promising avenue for surrogate modeling, but integrating data from multiple sources presents challenges.

Purpose of the Study:

• To develop and demonstrate multifidelity modeling approaches using neural network surrogates for cost-effective prediction of high-fidelity molecular properties.
• To integrate training data from multiple model forms and resolutions within quantum chemistry.
• To improve the efficiency and accuracy of potential energy surface predictions.

Main Methods:

• Utilized symmetry function-based atomic energy vectors as feature representations for molecular structures.
• Employed single-fidelity neural network training to map feature vectors to potential energy predictions.
• Implemented multifidelity topologies, including sequential and discrepancy-based formulations, to decompose high-fidelity mappings.

Main Results:

• Demonstrated methodologies on analytical test problems and applied them to predict potential energy for C5H5.
• Used high-fidelity B2PLYP-D3/6-311++G(d,p) and low-fidelity Hartree-Fock 6-31G data.
• Achieved an order of magnitude improvement in test error reduction or cost reduction for equivalent error, particularly with limited high-fidelity data.

Conclusions:

• Multifidelity neural network potential energy surface constructions are effective in enhancing prediction accuracy and reducing computational cost.
• The integration of data from multiple levels of theory significantly outperforms single-fidelity approaches.
• This approach offers a practical solution for situations with limited access to high-fidelity simulation data in computational chemistry.