Stanford CS153 Frontier Systems | Teaching AI to Touch Atoms

They expected to spend the first year doing purely computational “in silico” work, then later turn on a high-throughput autonomous lab. Instead, they found that building smaller semi-manual/semi-autonomous loops early was essential for steering the research program, selecting the right equipment to scale, and closing the experiment–model feedback loop fast.

Beyond predicting promising superconductors/semiconductors, the AI is valuable for mundane but consequential tasks: detecting sample mix-ups, controlling impurities, choosing mixing procedures, and interpreting characterization outputs (e.g., XRD patterns). These ‘local error corrections’ compound into real scientific throughput.

They argue that models won’t ‘think’ their way to new materials from textbooks alone; they must be grounded in experimental interaction. New data from synthesis and characterization is treated as core training signal to deepen a model’s physical understanding and improve future decisions.

Dorje claims Bayesian optimization often disappoints in practice because estimating uncertainty under distribution shift is even harder than point prediction. By contrast, active learning is critical in real-world settings where the model is competent in some regions and clueless in others; they run this style of iterative frontier-expansion daily via experiments.

They choose superconductors/semiconductors because both hinge on electron–atom interactions and interfaces, are central to modern technology, and face urgent materials bottlenecks (e.g., energy loss/heat in chips, scaling limits). The bet is that pushing AI hard on this ‘important direction’ beats hoping for magical generality.