Ångstrom used Claude Code to train a model that beat Meta's UMA-OMC (anycloud.sh)

Ångstrom AI, in collaboration with the University of Cambridge and AstraZeneca, has unveiled a groundbreaking machine learning model called CSP-MACE-Å, which dramatically accelerates crystal structure prediction (CSP) by achieving a 10,000x speedup over traditional density functional theory (DFT) calculations while maintaining comparable accuracy. This model outperformed Meta's UMA-OMC, the previous leading interatomic potential model for organic molecular crystals, thereby setting a new standard in the field. With CSP-MACE-Å, researchers can now conduct crystal structure analyses in minutes rather than weeks, significantly increasing the number of candidate structures examined and reducing risks associated with unexpected polymorphs during drug development. The significance of this advancement extends beyond computational efficiency; it directly impacts pharmaceutical development by enhancing the ability to predict potential crystal forms of drug compounds before they are released. The implications of CSP-MACE-Å stem from its integration with anycloud, a multi-cloud CLI that facilitates efficient GPU job management. Ångstrom utilized Claude Code to streamline the research experiment loop, enabling autonomous execution and monitoring of nearly 100,000 GPU jobs while implementing effective spend controls to manage cloud costs. This innovative framework positions Ångstrom as a leader in AI research, pushing the boundaries of what's possible in crystal structure prediction and drug formulation safety.