abstract
Machine learning (ML) based models have greatly enhanced the traditional materials discovery and design pipeline.
基于机器学习（ML）的模型极大完善了传统材料发现与设计流程。
Specifically, in recent years, surrogate ML models for material property prediction have demonstrated success in predicting discrete scalar-valued target properties to within reasonable accuracy of their DFT-computed values.
具体而言，近年来用于材料性能预测的机器学习代理模型，已成功实现离散标量型目标性能的预测，精度与密度泛函理论（DFT）计算值相符。
However, accurate prediction of spectral targets such as the electron Density of States (DOS) poses a much more challenging problem due to the complexity of the target, and the limited amount of available training data.
然而，电子态密度（DOS）这类谱型目标的精准预测，因目标本身复杂且可用训练数据有限，成为更具挑战性的问题。
In this study, we present an extension of the recently developed Atomistic Line Graph Neural Network (ALIGNN) to accurately predict DOS of a large set of material unit cell structures, trained to the publicly available JARVIS-DFT dataset.
本研究对新近提出的原子线图神经网络（ALIGNN）进行拓展，基于公开的 JARVIS‑DFT 数据集训练，精准预测大量材料晶胞结构的电子态密度。
Furthermore, we evaluate two methods of representation of the target quantity - a direct discretized spectrum, and a compressed low-dimensional representation obtained using an autoencoder.
此外，本研究评估了两种目标量表征方法：直接离散化谱表征，以及通过自编码器得到的压缩低维表征。
Through this work, we demonstrate the utility of graph-based featurization and modeling methods in the prediction of complex targets that depend on both chemistry and directional characteristics of material structures.
通过本研究，验证了基于图的特征化与建模方法，在预测同时依赖材料化学组成与结构方向特性的复杂目标时的有效性。
introduction
Although physics-based modeling approaches such as density-functional theory (DFT) have been used extensively in the development of material structure-property relationships, they generally incur a high computational cost.
尽管密度泛函理论（DFT）等基于物理的建模方法已被广泛应用于材料结构 - 性能关系研究，但其通常存在较高的计算成本。
As such, they are not well suited to materials discovery efforts that demand high-throughput screening to identify novel chemistries and material structures possessing certain desired combinations of physical and chemical properties.
因此，这类方法难以适用于需要高通量筛选、以发现具备目标物理化学性能组合的新型化学成分与材料结构的材料研发工作。
One of the primary potential applications of high-throughput screening with ML-based surrogate models is in the identification of materials that are well suited for use in energy storage applications.
基于机器学习代理模型的高通量筛选，其核心潜在应用之一是筛选适用于储能领域的材料。
Materials selected for such applications generally have a desired stability (i.e., target formation energy range), exhibit a particular conductive/insulating capability (i.e., target electronic bandgap range) and have a particular specific heat capacity (i.e., phononic and electronic heat capacity ranges).
这类应用所需材料通常需满足目标稳定性（即目标形成能区间）、特定导电 / 绝缘性能（即目标电子带隙区间）以及特定比热容（即声子与电子热容区间）。
Thus, a suitable workflow for identifying such materials is to shortlist a set of candidate materials (derived from the extensive materials search space) and perform DFT calculations for even more accurate property prediction on this shortlisted set of candidate materials.
因此，筛选此类材料的合理流程是：从庞大材料搜索空间中初选候选材料，再对候选材料进行 DFT 计算以实现更精准的性能预测。
ML-based surrogate models have been successfully employed in predicting scalar bulk material properties such as optical and electronic bandgaps 1, formation energy 2, Debye temperature 3, atomization energies 4 and polarizability of crystalline compounds 5.
基于机器学习的替代模型已成功应用于晶体材料标量体性能预测，包括光学带隙与电子带隙 1、形成能 2、德拜温度 3、原子化能 4 及极化率 5。
Of particular interest in this study is the electron Density of States (DOS), which reflects the number of states that may be occupied by the electrons in the material at a spectrum of energy levels.
本研究重点关注电子态密度（DOS），其表征材料中电子在不同能量区间可占据的量子态数目
The DOS can be used to compute directly many other useful physical properties of the material, such as the effective mass of electrons in charge carriers 10, the electronic contribution to heat capacity in metals 11, and the electronic bandgap of crystal structures.
电子态密度可直接计算材料多项关键物理性能，包括载流子电子有效质量 10、金属电子热容贡献 11 及晶体结构电子带隙。