|
| 1 | +--- |
| 2 | +jupytext: |
| 3 | + text_representation: |
| 4 | + extension: .md |
| 5 | + format_name: myst |
| 6 | + format_version: 0.13 |
| 7 | + jupytext_version: 1.16.1 |
| 8 | +kernelspec: |
| 9 | + display_name: Python 3 (ipykernel) |
| 10 | + language: python |
| 11 | + name: python3 |
| 12 | +--- |
| 13 | + |
| 14 | +# UMA Release Changelog |
| 15 | + |
| 16 | +This page documents the release history of UMA models, including new features, improvements, and bug fixes. |
| 17 | + |
| 18 | +--- |
| 19 | + |
| 20 | +## UMA 1.2 |
| 21 | + |
| 22 | +:::{admonition} Latest Release |
| 23 | +:class: tip |
| 24 | + |
| 25 | +**~50% faster, ~40% more accurate on Open Molecules test set, and expanded data coverage!** |
| 26 | +::: |
| 27 | + |
| 28 | +### Model Highlights |
| 29 | + |
| 30 | +::::{grid} 2 |
| 31 | +:gutter: 3 |
| 32 | + |
| 33 | +:::{grid-item-card} 🔋 Better Charge Handling |
| 34 | +Leading to improvements across molecular systems: |
| 35 | +- **~50%** relative improvement on biomolecules vs UMA-S-1.1 |
| 36 | +- **~30%** relative improvement on electrolytes vs UMA-S-1.1 |
| 37 | +::: |
| 38 | + |
| 39 | +:::{grid-item-card} 📊 Expanded Data Coverage |
| 40 | +Now **~520M total DFT calculations** including: |
| 41 | +- OC22 (oxide catalysts) |
| 42 | +- OC25 (electrolyte/inorganic interfaces) |
| 43 | +- Expanded OMol25 data including polymers (OPoly26) |
| 44 | +::: |
| 45 | + |
| 46 | +:::{grid-item-card} ⚡ Faster Inference |
| 47 | +**~50% speedup** for UMA-S with turbo mode compared to previous releases |
| 48 | +::: |
| 49 | + |
| 50 | +:::{grid-item-card} 🔧 Bug Fixes & Stability |
| 51 | +- Numerical and stability improvements for Hessians and phonons |
| 52 | +- More accurate diatomics and ionization potentials |
| 53 | +::: |
| 54 | + |
| 55 | +:::: |
| 56 | + |
| 57 | +### Available Tasks (7 total) |
| 58 | + |
| 59 | +| Task | Dataset | Domain | Status | |
| 60 | +|------|---------|--------|--------| |
| 61 | +| **oc20** | [OC20](https://arxiv.org/abs/2010.09990) | Heterogeneous Catalysis | No update | |
| 62 | +| **omat** | [OMat24](https://arxiv.org/abs/2410.12771) | Inorganic Materials | 📊 **Updated** — dimer data | |
| 63 | +| **omol** | [OMol-1.0](https://arxiv.org/abs/2505.08762) | Organic Molecules | 📊 **Updated** — polymers (OPoly26), ionization potentials, small molecule clusters, dimers | |
| 64 | +| **odac** | [ODAC23](https://arxiv.org/abs/2311.00341) | MOFs for Direct Air Capture | No update | |
| 65 | +| **omc** | [OMC25](https://arxiv.org/abs/2508.02651) | Molecular Crystals | No update | |
| 66 | +| **oc25** | [OC25](https://arxiv.org/abs/2509.17862) | Electrolyte/Inorganic Interfaces | 🆕 **Added to UMA** | |
| 67 | +| **oc22** | [OC22](https://arxiv.org/abs/2206.08917) | Oxide Catalysts | 🆕 **Added to UMA** | |
| 68 | + |
| 69 | +### Large-Scale Inference & Molecular Dynamics |
| 70 | + |
| 71 | +:::{note} |
| 72 | +UMA is built to easily scale up to multi-node, multi-GPU parallel inference with [Ray](https://www.ray.io/) under the hood. |
| 73 | +::: |
| 74 | + |
| 75 | +::::{grid} 2 |
| 76 | +:gutter: 3 |
| 77 | + |
| 78 | +:::{grid-item-card} 🚀 Large-Scale MD Simulations |
| 79 | +Run MD with ASE and LAMMPS on large-scale systems with ns/day speeds: |
| 80 | +- Battle tested up to **hundreds of GPUs** |
| 81 | +- Systems up to **1M atoms** |
| 82 | +- UMA handles all parallelism—no need to manually configure LAMMPS, Kokkos, MPI, CUDA, etc. |
| 83 | +::: |
| 84 | + |
| 85 | +:::{grid-item-card} 📦 Batched Simulations |
| 86 | +Client–server Ray framework enables batched simulations: |
| 87 | +- Structural relaxations |
| 88 | +- Molecular dynamics |
| 89 | +- **Up to 4x speedup** over serial calculations for batches of small systems on a single H100 GPU |
| 90 | +::: |
| 91 | + |
| 92 | +:::: |
| 93 | + |
| 94 | +--- |
| 95 | + |
| 96 | +## UMA 1.1 |
| 97 | + |
| 98 | +:::{admonition} Bug Fix Release |
| 99 | +:class: warning |
| 100 | + |
| 101 | +Fixed bug with size extensivity. |
| 102 | +::: |
| 103 | + |
| 104 | +### Available Tasks (5 total) |
| 105 | + |
| 106 | +| Task | Dataset | Domain | Status | |
| 107 | +|------|---------|--------|--------| |
| 108 | +| **oc20** | [OC20](https://arxiv.org/abs/2010.09990) | Heterogeneous Catalysis | No update | |
| 109 | +| **omat** | [OMat24](https://arxiv.org/abs/2410.12771) | Inorganic Materials | No update | |
| 110 | +| **omol** | [OMol25](https://arxiv.org/abs/2505.08762) | Organic Molecules | No update | |
| 111 | +| **odac** | [ODAC23](https://arxiv.org/abs/2311.00341) | MOFs for Direct Air Capture | No update | |
| 112 | +| **omc** | [OMC25](https://arxiv.org/abs/2508.02651) | Molecular Crystals | No update | |
| 113 | + |
| 114 | +--- |
| 115 | + |
| 116 | +## UMA 1.0 |
| 117 | + |
| 118 | +:::{admonition} Initial Release — May 2025 |
| 119 | +:class: note |
| 120 | + |
| 121 | +The first public release of UMA, introducing a unified model for atoms across multiple domains. |
| 122 | +::: |
| 123 | + |
| 124 | +### Available Tasks (5 total) |
| 125 | + |
| 126 | +| Task | Dataset | Domain | Status | |
| 127 | +|------|---------|--------|--------| |
| 128 | +| **oc20** | [OC20](https://arxiv.org/abs/2010.09990) | Heterogeneous Catalysis | 🆕 **New** | |
| 129 | +| **omat** | [OMat24](https://arxiv.org/abs/2410.12771) | Inorganic Materials | 🆕 **New** | |
| 130 | +| **omol** | [OMol25](https://arxiv.org/abs/2505.08762) | Organic Molecules | 🆕 **New** | |
| 131 | +| **odac** | [ODAC23](https://arxiv.org/abs/2311.00341) | MOFs for Direct Air Capture | 🆕 **New** | |
| 132 | +| **omc** | [OMC25](https://arxiv.org/abs/2508.02651) | Molecular Crystals | 🆕 **New** | |
| 133 | + |
| 134 | +### Key Features |
| 135 | + |
| 136 | +- **Mixture-of-Linear-Experts (MoLE)** architecture for high parameter count with fast inference |
| 137 | +- **Equivariant GNN** preserving energy conservation |
| 138 | +- **Multi-domain training** on 500M+ DFT calculations |
| 139 | +- **ASE Calculator integration** for seamless workflow adoption |
0 commit comments