Sparse Fine-tuning Boosts Materials AI Model Adaptation and Interpretability.

This research presents a novel fine-tuning approach designed to enhance the adaptation of pre-trained materials foundation models, also known as machine learning interatomic potentials. These models are crucial for approximating potential energy surfaces by leveraging general physicochemical knowledge. However, they often require specific calibration for diverse materials and to account for differences between training and application environments. The proposed method utilizes sparsity-promoting fine-tuning, which intelligently updates only a small subset of the model's parameters, sometimes as little as 0.5% to 3%. This selective updating process exploits the inherent structural properties of E(3)-equivariant materials foundation models. The technique has demonstrated performance on par with or superior to full fine-tuning and other adaptation methods for tasks like energy and force prediction across various molecular and crystalline systems. Beyond its efficiency, the method also shows strong generalizability, successfully applying to magnetic moment prediction and magnetism-aware total energy modeling. A significant advantage is its interpretability: analyzing the sparsity patterns reveals physically meaningful insights, such as increased d-orbital contributions in transition metal systems, making the model's adaptations more transparent.