Atomic Energy Accuracy of Neural Network Potentials: Harnessing Pretraining and Transfer Learning.

Area of Science:

• Computational Materials Science
• Machine Learning in Physics
• Quantum Chemistry

Background:

• Machine learning-based interatomic potentials (MLIPs) accurately predict potential energy surfaces (PESs).
• Atomic energy predictions within MLIPs are underexplored and their physical meaning is often disregarded.
• Inaccuracies in atomic energy predictions can compromise the robustness and transferability of Neural Network Potentials (NNPs).

Purpose of the Study:

• To investigate the impact of atomic energy prediction accuracy on the performance of Neural Network Potentials (NNPs).
• To demonstrate how error cancellation can mask inaccuracies in total energy predictions.
• To develop an improved method for training NNPs by focusing on atomic energy predictions.

Main Methods:

• Validation of atomic energy prediction inaccuracies using simulations of deformation and failure under tensile loading.
• Pretraining atomic energy predictions with empirical potentials.
• Applying transfer learning with density functional theory (DFT) data to enhance NNPs.

Main Results:

• Inaccuracies in atomic energy predictions were shown to reduce NNP robustness and transferability.
• Error cancellation in total energy predictions was identified as a masking effect for atomic energy errors.
• Pretraining and transfer learning significantly improved predictions of total energy, forces, and stress.

Conclusions:

• Atomic energy predictions are critical for developing high-quality and reliable MLIPs.
• Explicitly addressing atomic energy prediction accuracy enhances NNP performance.
• The proposed method improves the robustness and transferability of NNPs for materials simulations.