Learning Smooth and Expressive Interatomic Potentials for Physical Property Prediction
Authors: Xiang Fu, Brandon M Wood, Luis Barroso-Luque, Daniel S. Levine, Meng Gao, Misko Dzamba, C. Lawrence Zitnick
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The resulting model, e SEN, provides state-of-the-art results on a range of physical property prediction tasks, including materials stability prediction, thermal conductivity prediction, and phonon calculations. In this section, we evaluate e SEN in physical property prediction tasks: (1) materials stability prediction based on geometry optimization; (2) thermal conductivity prediction; and (3) phonon calculation. |
| Researcher Affiliation | Industry | 1Fundamental AI Research (FAIR) at Meta. Correspondence to: Xiang Fu <EMAIL>, C. Lawrence Zitnick <EMAIL>. |
| Pseudocode | No | The paper describes methods in prose and provides architecture diagrams (Figure 2) but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code is available at: https://github.com/facebookresearch/fairchem. Pretrained checkpoints are available at https://huggingface.co/facebook/OMAT24. |
| Open Datasets | Yes | We trained e SEN models under the same hyperparameters while ablating one design choice at a time. We construct out-of-distribution (OOD) MD simulation tasks for both inorganic materials and organic molecules using models trained on the MPTrj (Jain et al., 2013; Deng et al., 2023) and the SPICE-MACE-OFF (Eastman et al., 2023; Kov acs et al., 2023) datasets. The Matbench-Discovery benchmark evaluates a model s ability to predict ground-state (0 K) thermodynamic stability through geometry optimization and energy prediction. It is a widely used benchmark for evaluating ML models in materials discovery. |
| Dataset Splits | Yes | Since MPTrj lacks an official test split and various models are typically trained on distinct subsets of the data, we randomly selected 5000 samples from the subsampled Alexandria (s Alex) dataset (Schmidt et al., 2024; Barroso-Luque et al., 2024) for a fair comparison. The e SEN-30M-OAM model starts from the e SEN-30M-OMat model, and was finetuned for 1 epoch on a dataset constructed by combining the s Alex training dataset and 8 copies of the MPTrj training dataset. |
| Hardware Specification | Yes | We benchmarked the inference speed of our models against the similar sized MACE-OFF-L (Kov acs et al., 2023) (4.7M) on a single 80GB Nvidia A-100 GPU. |
| Software Dependencies | Yes | All models are benchmarked using the standard Python (3.12) runtime with Pytorch v2.4.0 (Paszke et al., 2019) and CUDA 12.1 (Nickolls et al., 2008). |
| Experiment Setup | Yes | Hyper-parameters used for model training are shown in Table 6. We train all models using a per-atom energy MAE loss, a force l2 loss, and a stress MAE loss. For direct-force models or direct-force pre-training, we use the same decomposed loss as described in Barroso-Luque et al. 2024. |