Transfer learning for chemically accurate interatomic neural network potentials.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Developing accurate interatomic potentials from ab initio methods is challenging.
• Machine learning offers a promising avenue for creating these potentials.

Purpose of the Study:

• To investigate transfer learning, specifically discriminative fine-tuning, for generating accurate interatomic neural network potentials.
• To assess the efficiency and limitations of transfer learning for organic molecules.

Main Methods:

• Utilized transfer learning (discriminative fine-tuning) on MD17 and ANI datasets.
• Pre-trained network parameters on density functional calculations and fine-tuned with energy labels.
• Investigated the impact of pre-training and fine-tuning dataset size and design.

Main Results:

• Pre-training significantly improves sample efficiency for models trained on ab initio data.
• Fine-tuning with energy labels alone can yield accurate atomic forces and enable large-scale simulations.
• Identified limitations related to dataset design and size in transfer learning.

Conclusions:

• Transfer learning, particularly discriminative fine-tuning, is effective for developing chemically accurate interatomic potentials for organic molecules.
• Pre-training enhances model efficiency, and energy-based fine-tuning is viable for simulations.
• Provided pre-trained and fine-tuned GM-NN potentials for broader application.