On 29 July, XtalPi launched XtalPi Science, a platform pairing large language models, scientific agents and large-scale robotic labs, alongside a Genius Agents matrix and a 27-member ‘Science Intelligence Open Ecosystem Alliance.’
XtalPi Science is billed as the world’s first AI-for-Science platform integrating LLMs, agents and automated experiment robots, packaged as on-demand ‘Science Tokens.’ Genius Agents orchestrate cross-disciplinary long-horizon tasks, cutting hallucination and closing a loop from digital hypothesis to physical verification.
The failure-data edge
XtalPi’s chief scientist Zhang Peiyu notes the platform’s rarest asset is failure data: 80 per cent of its 500,000 accumulated experiment records are negative samples absent from public literature. Trained on real physical data, its reaction-condition model SureRoute hits 4.6 per cent chemical hallucination, a sixth of frontier LLMs, with 51.7 per cent first-route accuracy, while SureRNX predicts failed experiments at 81 to 89 per cent, beating veteran chemists’ 38 to 60 per cent.
From lab to industry
The system compresses 5 to 10 trial iterations to an average 1.19, and already serves over 100 real drug and materials projects across 20-plus industrial scenarios, emitting 50,000 reaction-yield and 300,000 process records monthly. Physical AI, XtalPi argues, lands first where environments are structured and every experiment is valuable, the lab, before industry and home. It is moving an AI-discovered lung-cancer drug toward possible US approval in late 2026.
Read the original report (LeiPhone)
Translated and adapted from LeiPhone (leiphone.com).