Comparison of multifidelity machine learning models for potential energy surfaces

Multifidelity modeling is a technique for fusing the information from two or more datasets into one model. It is particularly advantageous when one dataset contains few accurate results and the other contains many less accurate results. Within the context of modeling potential energy surfaces, the low-fidelity dataset can be made up of a large number of inexpensive energy computations that provide adequate coverage of the N-dimensional space spanned by the molecular internal coordinates. The high-fidelity dataset can provide fewer but more accurate electronic energies for the molecule in question. Here, we compare the performance of several neural network-based approaches to multifidelity modeling. We show that the four methods (dual, Delta-learning, weight transfer, and Meng-Karniadakis neural networks) outperform a traditional implementation of a neural network, given the same amount of training data. We also show that the Delta-learning approach is the most practical and tends to provide the most accurate model.

Automated summary

Multifidelity modeling is a technique that combines information from multiple datasets into one model, which is advantageous when one dataset has few accurate results and the other has many less accurate results. In modeling potential energy surfaces, a large number of inexpensive energy computations can be used for the low-fidelity dataset, while the high-fidelity dataset provides fewer but more accurate electronic energies. In this study, neural network-based approaches for multifidelity modeling are compared, and it is found that the four methods mentioned outperform a traditional neural network with the same amount of training data. The Delta-learning approach is shown to be the most practical and provides the most accurate model.