Comparison of multifidelity machine learning models for potential energy surfaces.

Area of Science:

• Computational Chemistry
• Machine Learning
• Data Fusion

Background:

• Multifidelity modeling integrates datasets with varying accuracy and size.
• It's beneficial for modeling potential energy surfaces using computational chemistry data.
• Challenges exist in effectively combining low-fidelity (abundant, less accurate) and high-fidelity (scarce, accurate) data.

Purpose of the Study:

• To compare the performance of various neural network-based multifidelity modeling approaches.
• To evaluate their effectiveness in modeling potential energy surfaces.
• To identify the most practical and accurate method for this task.

Main Methods:

• Implementation and comparison of four neural network multifidelity methods: dual, Δ-learning, weight transfer, and Meng-Karniadakis networks.
• Training and testing these methods using datasets of low-fidelity (many inexpensive energy computations) and high-fidelity (fewer accurate electronic energies).
• Benchmarking against a traditional neural network implementation with identical training data.

Main Results:

• All four multifidelity neural network approaches outperformed traditional neural networks.
• The Δ-learning approach demonstrated superior practicality and accuracy among the tested methods.
• The study confirms the advantage of multifidelity modeling for potential energy surface calculations.

Conclusions:

• Neural network-based multifidelity modeling is a powerful technique for enhancing model accuracy and efficiency.
• The Δ-learning method stands out as a highly effective and practical approach for potential energy surface modeling.
• This work provides valuable insights for computational chemists and machine learning practitioners.