Machine learning(ML)has accelerated the process of materials classification,particularly with crystal graph neural network(CGNN)architectures.However,advanced deep networks have hitherto proved challenging to build an...Machine learning(ML)has accelerated the process of materials classification,particularly with crystal graph neural network(CGNN)architectures.However,advanced deep networks have hitherto proved challenging to build and train for quantum materials classification and property prediction.We show that faithful representations,which directly represent crystal structure and symmetry,both refine current ML and effectively implement advanced deep networks to accurately predict these materials and optimize their properties.Our new models reveal the previously hidden power of novel convolutional and pure attentional approaches to represent atomic connectivity and achieve strong performance in predicting topological properties,magnetic properties,and formation energies.With faithful representations,the state-of-the-art CGNN accurately predicts quantum chemistry materials and properties,accelerating the design and discovery and improving the implicit understanding of complex crystal structures and symmetries.On two separate benchmarks,our non-graphical neural networks achieve near parity with the CGNN architecture,making them viable alternatives.展开更多
基金supported as part of the Center for Energy Efficient Magnonics, an Energy Frontier Research Center funded by the U.S. Department of Energy, Office of Science, Basic Energy Sciences, under Award number DE-AC02-76SF00515.
文摘Machine learning(ML)has accelerated the process of materials classification,particularly with crystal graph neural network(CGNN)architectures.However,advanced deep networks have hitherto proved challenging to build and train for quantum materials classification and property prediction.We show that faithful representations,which directly represent crystal structure and symmetry,both refine current ML and effectively implement advanced deep networks to accurately predict these materials and optimize their properties.Our new models reveal the previously hidden power of novel convolutional and pure attentional approaches to represent atomic connectivity and achieve strong performance in predicting topological properties,magnetic properties,and formation energies.With faithful representations,the state-of-the-art CGNN accurately predicts quantum chemistry materials and properties,accelerating the design and discovery and improving the implicit understanding of complex crystal structures and symmetries.On two separate benchmarks,our non-graphical neural networks achieve near parity with the CGNN architecture,making them viable alternatives.
