Foundation Models
A foundation model is a large machine learning model pre-trained on broad, diverse data and then adapted (fine-tuned) with relatively little task-specific data to many downstream tasks. In chemistry and materials science, examples include universal interatomic potentials trained on large DFT datasets and models that combine several data types, such as structures, spectra, images and text, in one shared representation.
Multimodal foundation models for heterogeneous catalysis (2026)
The roadmap by Xin, Kitchin, López, Schweitzer and co-authors makes multimodal foundation models the second pillar of AI in catalysis, after an AI-ready data ecosystem (Xin et al. 2026). It describes how such models would combine representation learning, transfer learning, generative modelling, contrastive alignment of modalities and reinforcement learning. It cites examples such as a graph-text contrastive model that reduced adsorption energy error by about 10% compared with a text-only model, and transfer of reactivity knowledge with as few as ten new experimental points.
The authors argue that these models are still fragile: they degrade under domain shift, lack calibrated uncertainty and are weakly grounded in physics. They call for physics-informed inductive biases, uncertainty quantification, out-of-distribution detection, standardized model and data cards, multimodal benchmarks and time-split validation, and they want predictions to come with explanations that scientists can test. Summary: xin2026.