In this video I highlight some results from a recent paper titled "Investigating the Error Imbalance of Large-Scale Machine Learning Potentials in Catalysis" ([ Ссылка ]). We investigated why it appears that the accuracy of large-scale MLPs has plateaued in recent years. We show it is not likely due to convergence errors, but instead due to bad labels in the data. We show some paths to improve the accuracy, and tradeoffs in the choices that get made.
