Accuracy and Efficiency Benchmarks of Pretrained Machine Learning Potentials for Molecular Simulations.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Pretrained Machine Learning Interatomic Potentials (MLIPs) are rapidly advancing.
• Selecting the optimal MLIP for specific applications is difficult due to the wide variety of available models.

Purpose of the Study:

• To benchmark 15 pretrained MLIPs.
• To provide an objective comparison of MLIPs based on accuracy, speed, memory usage, and simulation stability.
• To identify factors influencing MLIP accuracy and performance.

Main Methods:

• Benchmarking of 15 distinct pretrained MLIPs.
• Evaluation metrics included accuracy, computational speed, memory requirements, and simulation stability.
• Analysis of the correlation between model characteristics (parameters, training set size, energy terms) and performance.

Main Results:

• Accuracy is strongly correlated with the number of model parameters and the size of the training dataset.
• Explicit Coulomb energy terms did not improve accuracy in this study.
• Computational speed and memory usage depend significantly on the MLIP architecture, not just model size.

Conclusions:

• The study provides a quantitative basis for selecting appropriate MLIPs for molecular simulations.
• Model complexity and training data volume are critical for achieving high accuracy.
• MLIP architecture is a crucial factor for computational efficiency.