AI models for science are models developed to address AI for science. They include foundation models for science but need not be a foundation model; for example, CosmoFlow is a CNN-based model developed to solve one specific problem in cosmology but is not transferable to solving other cosmological problems.
Benchmarks
MLPerf HPC includes:
- DeepCAM: CNN-based climate segmentation model
- CosmoFlow: Predicts cosmological parameters?
- DimeNet++: Quantum molecular model?
- AlphaFold2: Protein folding model
The UK BASE-II AI benchmarks include:
- Scientific models
- Numerical Weather Prediction (NWP) tutorial model: Forecasting model(?)
- DeepCAM: CNN-based climate segmentation model
- EquiformerV2: GNN-based molecular structure model
- Non-science-specific models
- GPT-2 (text transformer)
- BERT
- T5 Model
- ResNet (CNN)
- OLMo (text transformer)
- Megatron(text transformer)