Scholarlink

<p>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 <i>faithful representations</i>, 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.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Faithful novel machine learning for predicting quantum properties

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年7月26日

查看全文
  • Gavin Nop,
  • Micah Mundy,
  • Jonathan D. H. Smith,
  • Durga Paudyal

摘要

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.