Training Machine Learning Interatomic Potentials on Cloud GPUs
· 13 min read
Machine learning interatomic potentials (MLIPs) are replacing classical force fields for many use cases in computational chemistry — and made simulations that used to take weeks of DFT compute possible in minutes on a single GPU. This post is a practical guide to training your own MLIP on cloud GPU spot instances: which architecture to pick, what it actually costs, and how to keep training going through preemption.