ALCF resources enable researchers to train, fine-tune, and optimize AI models for scientific research. From foundation models and large language models to surrogate models and domain-specific neural networks, users can leverage ALCF systems to tackle training workloads that exceed the capabilities of conventional computing environments.
By combining advanced computing resources, scalable software environments, and staff expertise, ALCF helps researchers accelerate experimentation and explore larger model and parameter spaces.
Researchers can use ALCF systems for a variety of tasks, including:
- Training foundation models and large language models
- Fine-tuning existing models for scientific applications
- Developing surrogate models for complex simulations
- Performing hyperparameter optimization and architecture searches
- Building multimodal models with simulation and experimental data
- Scaling training across large-scale GPU and accelerator systems
Applications
Examples of AI model development projects that leveraged ALCF systems include:
- Foundation models for battery materials discovery
- Multimodal frameworks for protein design
- Domain-specific language models for materials science
- A diffusion model for Earth system prediction
Systems
Researchers can train and fine-tune AI models using ALCF's leadership-class computing systems, including Aurora, Polaris, and the ALCF AI Testbed. Together, these resources provide access to large-scale GPU platforms and emerging AI accelerators, enabling researchers to explore a wide range of model architectures and training approaches.
