Adversarial approaches,which intentionally challenge machine learning models by generating difficult examples,are increasingly being adopted to improve machine learning interatomic potentials(MLIPs).While already prov...Adversarial approaches,which intentionally challenge machine learning models by generating difficult examples,are increasingly being adopted to improve machine learning interatomic potentials(MLIPs).While already providing great practical value,little is known about the actual prediction errors of MLIPs on adversarial structures and whether these errors can be controlled.We propose the Calibrated Adversarial Geometry Optimization(CAGO)algorithm to discover adversarial structures with userassigned errors.Through uncertainty calibration,the estimated uncertainty of MLIPs is unified with real errors.By performing geometry optimization for calibrated uncertainty,we reach adversarial structures with the user-assigned target MLIP prediction error.Integrating with active learning pipelines,we benchmark CAGO,demonstrating stable MLIPs that systematically converge structural,dynamical,and thermodynamical properties for liquid water and water adsorption in a metal-organic framework within only hundreds of training structures,where previously many thousands were typically required.展开更多
基金supported by the Research Council of Norway through the Centre of Excellence Hylleraas Centre for Quantum Molecular Sciences(Grant 262695)the Young Researcher Talent grants 344993 and 354100+2 种基金We acknowledge the EuroHPC Joint Undertaking for awarding this project access to the EuroHPC supercomputer LUMI,hosted by CSC(Finland)and the LUMI consortium through a EuroHPC Regular Access call(Grants EHPC-REG-2023R02-088,EHPC-REG-2023R03-146)Support was also received from the Centre for Advanced Study in Oslo,Norway,which funded and hosted the SLB Young CAS Fellow research project during the academic year of 23/24 and 24/25Part of the simulations were performed on resources provided by Sigma2—the Norwegian National Infrastructure for High-Performance Computing and Data Storage(grant numbers NN4654K and NS4654K).
文摘Adversarial approaches,which intentionally challenge machine learning models by generating difficult examples,are increasingly being adopted to improve machine learning interatomic potentials(MLIPs).While already providing great practical value,little is known about the actual prediction errors of MLIPs on adversarial structures and whether these errors can be controlled.We propose the Calibrated Adversarial Geometry Optimization(CAGO)algorithm to discover adversarial structures with userassigned errors.Through uncertainty calibration,the estimated uncertainty of MLIPs is unified with real errors.By performing geometry optimization for calibrated uncertainty,we reach adversarial structures with the user-assigned target MLIP prediction error.Integrating with active learning pipelines,we benchmark CAGO,demonstrating stable MLIPs that systematically converge structural,dynamical,and thermodynamical properties for liquid water and water adsorption in a metal-organic framework within only hundreds of training structures,where previously many thousands were typically required.
