Foundation Models for Atoms
Quantum-mechanical methods such as density functional theory (DFT) are accurate but limited to small systems and short timescales, while classical force fields are fast but often not accurate or transferable enough. Machine-learned interatomic potentials (MLIPs) could break this trade-off between accuracy and scale, and the field is now moving at remarkable speed. So-called universal or "foundation" models (e.g. MACE-MP, UMA, MatterSim, Orb, the DPA/OpenLAM series), pre-trained on tens to hundreds of millions of DFT calculations spanning the periodic table, approach DFT-level accuracy at a small fraction of the cost. They can be applied out-of-the-box ("zero-shot") to almost any chemistry, then fine-tuned to a specific system with a small amount of targeted data, shortening the path from question to result for both academic and industrial users.
Practical information:
Georganiseerd door:
In this webinar, we will give a conceptual tour of this rapidly evolving landscape and what it means for everyday computational research on HPC systems. We will cover:
- what universal MLIPs are and how they differ from classical force fields and system-specific ML potentials;
- the practical pre-train → fine-tune workflow and when out-of-the-box use is (and is not) good enough;
- the new generation of batched, GPU-accelerated simulation engines (e.g. TorchSim, kUPS) built for high-throughput MLIP simulation;
- how to choose and trust a model using open benchmarks;
- and close with a brief outlook on where the field is heading, including ML making its way into electronic-structure theory itself, with learned density functionals, and a short demo on European HPC resources. Throughout, we will point to open models, datasets, benchmarks, and codes that participants can try on their own problems right after the webinar.