Fujitsu and Tohoku apply causal AI to unlock superconductivity insights

Fujitsu and Tohoku’s work is a concrete example of AI moving from “pattern finder” to “hypothesis generator” in hard science. By using a causal discovery approach on dense ARPES spectroscopy data, they reduced an intractably large causal graph to something physicists could actually interpret, and then tied that back to a specific superconducting mechanism in a Kagome material. That’s qualitatively different from just fitting a regression or clustering spectra; it’s AI helping decide what underlying story best explains the data.

From an AGI perspective, this is exactly the sort of loop we expect to see more: AI proposes plausible causal structures, humans validate them, and the combined system explores exponentially larger hypothesis spaces than unaided scientists could. As tools like Fujitsu’s Kozuchi platform mature and become more widely accessible, they’ll accelerate scientific discovery across materials, energy, and biology—areas where new knowledge often feeds back into better compute, sensors, and devices for AI itself.

Strategically, Fujitsu is also positioning itself as a player in “AI for science,” not just enterprise IT. That pits it against the likes of Google DeepMind, Microsoft, and various national labs that are also trying to prove AI can deliver real scientific breakthroughs, not just better chatbots. Whoever builds the best AI copilots for scientists will wield outsized soft power in the next phase of the AGI race.