Photographer: Samuel Avraham.

The event page includes the full programme, and talks were recorded and are available on demand. 

A clear through-line for EDUCATE and OKC attendees was the move toward physically informed, interpretable machine learning: models that not only make accurate predictions but also help scientists learn about the underlying systems and quantify uncertainty. For newcomers to ML, the sessions underscored just how broad today’s toolbox is—far beyond the few model classes dominating headlines—and how important it is to match model structure to the scientific question. “Seeing that machine-learning models don’t have to be black boxes was very useful,” said Alex Brown (OKC).

Discussions repeatedly showed that methods travel well across fields while data differ. Many domains now face the same challenge: sifting ever larger, noisier datasets to make genuine progress. Several talks highlighted that shared landscape, including explainable AI in cosmology, data-driven gravitational lensing, neural operators for scientific simulation, and AI-enabled weather forecasting—with contributions spanning NVIDIA (Jean Kossaifi), Google DeepMind (Ilan Price), SandboxAQ (Stefan Leichenauer), CuspAI (Max Welling), and universities across Europe and North America. 

A recurring theme was science’s shift from experiment-first, to data-driven, to in-silico experimentation with digital twins—a trajectory now visible across biology, climate, and the physical sciences. That framing connected naturally to talks on materials and climate, where AI-guided materials discovery (e.g., efforts to design sorbents for CO₂ capture) was presented as a concrete near-term impact area. 

Participants also engaged candidly with compute and sustainability. Rather than treating energy use as a rebuttal to AI, several discussions emphasized designing models and pipelines that reduce overall cost by replacing expensive brute-force simulations, improving forecasting lead times, or accelerating search in high-dimensional design spaces—net improvements that matter at scale. In parallel, the industry–academia comparison surfaced complementary strengths: rapid iteration and performance tuning on one side; interpretability, physical consistency, and open science on the other. The format made those differences productive, and suggestions for future editions included more small-group time to accelerate collaborations. 

“We left with a clearer view of how AI-focused companies and academia can work together,” said Ariel Goobar. “The focused discussions are likely to translate into new collaborations for EDUCATE and open fresh research directions.”