Optimizing the architecture of Behler-Parrinello neural network potentials

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Neural network potentials (NNPs) are crucial for molecular simulations.
• NNP architecture is often fixed post-initialization, potentially limiting accuracy.
• The impact of training set size on NNP architecture and accuracy is not fully understood.

Purpose of the Study:

• To investigate the effect of training set size on Behler-Parrinello neural network potential accuracy.
• To determine optimal neural network architectures for varying dataset sizes.
• To analyze the influence of architectural choices on NNP performance.

Main Methods:

• Trained Behler-Parrinello neural network potentials on QM9 and 3BPA datasets.
• Varied training set sizes and corresponding network architectures.
• Analyzed the impact of descriptor complexity, network depth (number of hidden layers), and activation functions.

Main Results:

• Adjusting network architecture to match training set size significantly improved NNP accuracy.
• Both insufficient and excessive numbers of fitting parameters negatively impacted accuracy.
• Two hidden layers and unbounded activation functions provided the best performance for the studied NNPs.

Conclusions:

• Neural network potential architecture should be adapted to the training set size for optimal accuracy.
• Careful consideration of fitting parameters, network depth, and activation functions is essential for developing accurate NNPs.
• This study provides guidelines for designing more accurate and efficient neural network potentials.