We employ machine learning to derive tight-binding parametrizations for the electronic structure of defects. We test several machine learning methods that map the atomic and electronic structure of a defect onto a sparse tight-binding parameterization. Since Multi-layer perceptrons (i.e., feed-forward neural networks) perform best we adopt them for our further investigations. We demonstrate the accuracy of our parameterizations for a range of important electronic structure properties such as band structure, local density of states, transport and level spacing simulations for two common defects in single layer graphene. Our machine learning approach achieves results comparable to maximally localized Wannier functions (i.e., DFT accuracy) without prior knowledge about the electronic structure of the defects while also allowing for a reduced interaction range which substantially reduces calculation time. It is general and can be applied to a wide range of other materials, enabling accurate large-scale simulations of material properties in the presence of different defects.

我们使用机器学习来为缺陷的电子结构推导出紧密绑定的参数化。我们测试了几种机器学习方法，这些方法将缺陷的原子和电子结构映射到稀疏紧束缚参数化上。由于多层感知器（即前馈神经网络）表现最好，我们采用它们进行进一步的研究。我们证明了我们对一系列重要电子结构特性的参数化的准确性，例如单层石墨烯中两种常见缺陷的能带结构、局部状态密度、传输和能级间距模拟。我们的机器学习方法取得了与最大局部化 Wannier 函数相当的结果（即，DFT 精度）无需事先了解缺陷的电子结构，同时还允许减小相互作用范围，从而显着减少计算时间。它是通用的，可以应用于广泛的其他材料，能够在存在不同缺陷的情况下对材料特性进行精确的大规模模拟。