Sintered neodymium-iron-boron(NdFeB)magnets are indispensable in high-performance applications,but their optimization is challenged by complex structure-property relationships and limited data.In this work,we curate t...Sintered neodymium-iron-boron(NdFeB)magnets are indispensable in high-performance applications,but their optimization is challenged by complex structure-property relationships and limited data.In this work,we curate the first multi-domain database for this system(1994 industrial and academic samples)and systematically evaluate active learning(AL)strategies on classical and quantum-enhanced regressors.First,our“domain-aware”analysis reveals quantitative differences in design heuristics between industrial and academic data.Second,we present a methodological blueprint for integrating quantum kernel regression into an AL framework using a bootstrapped ensemble for uncertainty quantification.Finally,and most significantly,our results reveal AL effectiveness is strongly model-dependent.Its advantage ranges from significant acceleration(Random Forest,SVR)to being diminished(XGBoost),or even inverted—proving detrimental compared to random sampling—as shown in our quantum-enhanced SVR case study.This finding provides critical new insights for the strategic application of machine learning in materials discovery.展开更多
基金supported by National Key Research & Development Program of China (No. 2024YFC3906900)National Natural Science Foundation of China (No. 22173114, 22333003 to Y. Ma+3 种基金52401251 to H. Xu)Strategic Priority Research Program (XDB0500101)Youth Innovation Promotion Association (No. 2022168)Project of Ganjiang Innovation Academy (E455F001) of Chinese Academy of Sciences. Some of the computational experiments were implemented in the ORISE and ERA supercomputers, we are also highly appreciated the helps from the supporting team.
文摘Sintered neodymium-iron-boron(NdFeB)magnets are indispensable in high-performance applications,but their optimization is challenged by complex structure-property relationships and limited data.In this work,we curate the first multi-domain database for this system(1994 industrial and academic samples)and systematically evaluate active learning(AL)strategies on classical and quantum-enhanced regressors.First,our“domain-aware”analysis reveals quantitative differences in design heuristics between industrial and academic data.Second,we present a methodological blueprint for integrating quantum kernel regression into an AL framework using a bootstrapped ensemble for uncertainty quantification.Finally,and most significantly,our results reveal AL effectiveness is strongly model-dependent.Its advantage ranges from significant acceleration(Random Forest,SVR)to being diminished(XGBoost),or even inverted—proving detrimental compared to random sampling—as shown in our quantum-enhanced SVR case study.This finding provides critical new insights for the strategic application of machine learning in materials discovery.
