Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
UMA: A Family of Universal Models for Atoms
Authors: Brandon M. Wood, Misko Dzamba, Xiang Fu, Meng Gao, Muhammed Shuaibi, Luis Barroso-Luque, Kareem Abdelmaqsoud, Vahe Gharakhanyan, John Kitchin, Daniel S Levine, Kyle Michel, Anuroop Sriram, Taco S Cohen, Abhishek Das, Sushree Sahoo, Ammar Rizvi, Zachary Ulissi, Larry Zitnick
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate UMA models on a diverse set of applications across multiple domains and find that, remarkably, a single model without any fine-tuning can perform similarly or better than specialized models. Our evaluation contains two main components: (1) a diverse set of held-out test splits (Table 2) and (2) a suite of practically important benchmarks (Table 3). |
| Researcher Affiliation | Collaboration | 1FAIR at Meta 2CMU |
| Pseudocode | No | The paper describes the architecture and training procedures in detail in text and with figures, but it does not include explicitly labeled pseudocode or algorithm blocks. The training procedure is outlined in Section 2.4 and detailed further in Appendix A. |
| Open Source Code | Yes | We are releasing the UMA code, weights, and associated data to accelerate computational workflows and enable the community to build increasingly capable AI models. All code and data used for training UMA is publicly available. |
| Open Datasets | Yes | The Open Molecules 2025 (OMol25) [43] and Open Materials 2024 (OMat24) [5] datasets cover both molecules and materials respectively. The Open Catalyst 2020 (OC20) [11] and Open DAC 2025 (ODAC25) [68] datasets are useful for modeling the interactions of molecules and materials. Finally, the Open Molecular Crystals 2025 (OMC25) [23] dataset specializes in modeling the interaction between molecules in periodic structures. |
| Dataset Splits | Yes | Our evaluation contains two main components: (1) a diverse set of held-out test splits (Table 2) and (2) a suite of practically important benchmarks (Table 3). In Table 2, we report the test-set performance on two OOD test sets: WBM [74] and high entropy alloy (HEA). The validation and test results are shown in Tables 16 and 17, respectively. |
| Hardware Specification | Yes | Table 1: Inference speed and max atoms measured on Nvidia H100 with a periodic system that has 50 neighbors per atom within 6Å, see Appendix D. Table 7: Training Times for UMA models. Model Stage GPUs in Parallel Training Days GPU-Type. UMA-S Direct Pre-train 128 5 H200 140GB. UMA-S Conserve Fine-tune 256 5 H200 140GB. UMA-M Direct Pre-train 128 14 H200 140GB. UMA-M Conserve Fine-tune 256 14 H200 140GB. UMA-L Direct Pre-train 128 25 H100 80GB. UMA-L Stress Fine-tune 128 4 H100 80GB. UMA-L FP32 Fine-tune 128 2 H100 80GB. |
| Software Dependencies | Yes | For fair comparisons against other models, we used pytorch2.6.0, cuda12.4, python3.12 and TF-32 precision universally on a H100 80GB GPU. |
| Experiment Setup | Yes | Appendix A: Training Details and Hyperparameters. Table 4: Summary of main training-related hyper-parameters for the pre-training and fine-tuning stages. These hyper-parameters are shared among model sizes. Table 5: Hyper-parameters for UMA models of different sizes. |