机器学习在热电材料中的应用

MACHINE LEARNING APPLICATION IN THERMOELECTRIC MATERIALS

房亮

11749111

硕士

材料加工工程

刘玮书

2019-06-01

2019-07-10

哈尔滨工业大学

深圳

热电材料是一种可以实现热能和电能直接转换的功能材料，在航空航天中的原子能电池，电子通讯领域的固态制冷，汽车尾气废热发电，以及物联网自供能等领域都有重要的应用价值。新型热电材料的发现对推动热电转换技术的进步有重要意义。回顾热电材料的发展历程，最早可以追溯到Seebeck通过试错法所做出的探索性研究。上世纪50年代，凝聚态物理理论的应用发现了经典的Bi2Te3和PbTe等热电材料。近年来，第一性原理计算的发展也极大地促进了如Mg3Sb2等新型热电材料的发现。然而，与已知的化合物相比，未知化合物的数量更为庞大，本论文借助机器学习的方法来筛选还没有确定结构的未知化合物中潜在的热电材料。根据经典的Bi2Te3材料，我们利用等价电子元素取代构建了一个含有720种化合物的M2X3型热电材料库，其中在ICSD数据库中发现了80种具有晶体结构的化合物。我们选用组成元素的物理性质（例如原子尺寸，电负性，密度等）来定义具有通式M1M2X1X2X3 (M1+M2: X1+X2+X3= 2: 3) 的化合物的特征，并利用随机森林+贝叶斯优化方法构建机器学习预测模型。通过模型学习已知结构的80种化合物，获取结构分类的规律，并进一步用来预测未知化合物。研究发现由于用于生成规则的学习样本太少，对于完整预测七大晶系的多任务预测准确度不高。因此，本文进一步修改策略，仅通过已知样本获取是否是与Bi2Te3材料一致的三方结构。基于机器学习的交叉验证结果显示，仅判断一个化合物是否是三方结构化合物的单任务预测模型的准确率可以高达0.94。利用参考文献中10种已知结构M2X3型化合物数据进一步验证模型，可实现0.90的准确度。因此，本文最终使用该模型来筛选材料库中具有类似三方结构的新化合物，并获得70+个潜在和Bi2Te3结构相近的新化合物。本文通过特征选择和探索性数据分析发现了影响M2X3型热电材料结构的敏感参数，进一步提出基于电负性、离子半径、熔点等特征构成的规则，此规则也能获得与机器学习模型相当的准确度。然而，仅具有结构相似，尚无法说明该材料就一定是一个好的热电材料。禁带宽度是人们判断一个化合物是否是一个潜在热电材料的重要指标。本文基于Slack关于禁带宽度与阳离子(M)和阴离子(X)之间电负性差值绝对值的经验关系，提出了双重相似性的标准，即将结构相似和阴阳离子平均电负性差相似作为新型热电材料筛选的判决指标，并发现有潜力的新型热电材料如Sb2TeSeS、Sb2S2Te、SbFeTe3等。

Thermoelectric material is a functional material that can directly convert thermal energy and electrical energy, and has been used in atomic energy battery in aerospace and refrigeration in electronic communications, and also shows a promising in waste heat harvesting and self-powered energy supply in IoT (Internet of Things). The continuously discovering of new materials has made a significant contribution to the advance in thermoelectric field. The discovering of new thermoelectric materials can be traced back to Seebeck's exploratory research through trial-and-error efforts. In the 1950s, classical thermoelectric materials such as Bi2Te3 and PbTe were found due to the progress in condensed matter physics. Recent years, the development of first principle calculation has greatly promoted the discovery of new thermoelectric materials such as Mg3Sb2. However, compared with known-structure materials, the number of unknown-structure materials is even bigger. Here, we proposed a machine learning assistant discovering of the new thermoelectric materials. According to the classic Bi2Te3 material, we constructed a M2X3-type thermoelectric material library with 720 compounds using equivalent valence electron substitution in which 80 compounds were found to have crystalline structures in the ICSD database. The physical properties of constituent elements (such as atomic size, electronegativity, density, etc.) were used to define the feature of the compounds with a general formula M1M2X1X2X3 (M1+M2:X1+X2+X3=2:3). A random forest plus bayesian optimization of hyperpatameters was used as the machine leaning method, and the 80 compounds with known structures were used as the foundation database to learn the structure-classification rules and predict new materials. The first objective is to find the rule that could identify the compounds with the same rhombohedral structure of Bi2Te3. The first try found that the number of learning samples used to generate the rules was too small, and the accuracy of multi-task prediction for the complete prediction of the seven major crystal systems was rather limited. Therefore, we modified the strategy to only identify the compounds with the same rhombohedral structure of Bi2Te3. The cross validation of the machine learning process showed a high accuracy of 0.94 for the prediction of rhombohedral compounds. The prediction of 10 M2X3-type compounds that were found in the recent references, showed an accuracy of 0.9. We finally used the code to predict new compounds with similar rhombohedral structure and obtained 70+ new compounds that could have similar structure with Bi2Te3. Furthermore, it was found the important features affecting the structure of M2X3–type thermoelectric materials through feature selection and exploratory data analysis. So we proposed structure rules based on elemental features such as electronegativity, ionic radius, and melting point. This rule could achieve the same accuracy as the machine learning model. However, only the similar structure can not guarantee the candidate must be a good thermoelectric material. The band gap is an important indicator to judge whether a compound is a potential thermoelectric material. Furthermore, based on the Slack’s early work on the relationship between the band gap and average electronegativity difference between the cations (M) and anions (X), we proposed a criteria of dual-similarity, i.e. structure similarity and electronegativity similarity, and some potential new compounds were given out, such as Sb2S2Te, SbFeTe3, Sb2TeSeS.

热电材料 机器学习 随机森林 贝叶斯优化 晶体结构预测

thermoelectric material machine learning random forest bayesian optimization crystal structure prediction

中文

联合培养

南方科技大学

房亮. 机器学习在热电材料中的应用[D]. 深圳. 哈尔滨工业大学,2019.