6-14. Machine learning and big data towards ultra-high temperature

Shenghong Ju

Sino-British international low-carbon college, Shanghai jiao tong university/ Materials genome joint research center/ Shool of science and engineering/ University of Tokyo

Abstract: Designing materials with ultimate high/low thermal conductivity holds its critical importance in applications of heat exchanger, heat radiators, thermal interface materials, thermoelectrics, thermal barrier coating and insulators. However, two bottlenecks limit the designing efficiency: material selection and nanostructure designing. The traditional material design method is limited by the various parameters as well as their coupling effects, leading to high cost and low efficiency. In recent years, the big data and artificial intelligence have been rapidly developed and popularized. If we can combine the rapid development of informatics with the traditional material calculation/experiment, it will help to accelerate the material design. As the fourth paradigm of research in addition to theory, calculation and experiment, materials informatics integrates the data science and material calculation/experiment and is attracting more and more researchers' attention due to its unique advantages in materials design. In the past few decades, materials informatics has been successfully applied to the design of cathode materials for lithium-ion batteries, drugs, polymers, catalysts, but the development in heat transfer field is still in its infancy.

In this work, I will introduce two recent successful applications of materials informatics in the design of thermal functional materials: (1) High throughput screening of ultimate high/low thermal conductivity materials. By using transfer learning and harmonic phonon property, the prediction model for thermal conductivity has been successfully built up based on small amount of training data. The trained model is employed to screen over 60000 compounds in materials project database, and series of diamond-like materials with high thermal conductivity have been found, which provides more choices and possibilities for the design of high thermal conductivity materials and devices. (2) Nanostructure designing using Bayesian optimization and Monte Carlo tree search. The Bayesian optimization and Monte Carlo tree search algorithms are combined with the traditional heat transport calculation to realize the intelligent structure design of nanostructures for high/low thermal conductivity. This method is universal and can be extended to a variety of complex energy transport design systems, including thermal radiation, photons, magnetons, etc. Those successful applications have shown the high efficiency of materials informatics for materials design and optimization, which can effectively reduce the development cycle and cost for new materials.

Keywords: Materials informatics; Thermal functional materials; Database screening; Nanostructure design and optimization

基于材料信息学的热功能材料设计与优化

鞠生宏^*

上海交通大学中英国际低碳学院/材料基因组联合研究中心/材料科学与工程学院，东京大学

摘要：开发设计具有极限导热特性的热功能材料在换热器、热界面、热电、热涂层、热绝缘材料等领域具有很大的应用前景，然而材料选择和结构设计这两个瓶颈问题限制了导热材料设计的效率及发展。传统设计方法受限于材料设计过程中的诸多参数选择及耦合效应，实现最佳性能的材料设计往往效率低、成本高。近年来，大数据及人工智能作为新兴学科得到了快速发展和推广，如能将快速发展的信息学和传统材料特性研究方法结合，将有助于解决以往设计难、效率低的困境。基于这个思路，材料信息学方法应运而生。材料信息学作为除理论、计算、实验之外第四维的研究工具，融合了数据科学以及材料性能计算、实验等方法，在材料研发与设计中具有独特的优势，正吸引越来越多研究人员的广泛关注。在过去的几十年中，材料信息学已经被成功应用于设计锂离子电池的正极材料、药物、聚合物、催化剂等领域，但在传热领域的发展还处于起步阶段。

本文将介绍最近材料信息学在热功能材料设计方面成功的两个应用：（1）基于高通量计算的高热导率材料搜索，利用转移学习的思路和快速的晶体简谐声子特性计算，基于少量的训练数据成功构建了有效的热导率预测模型，搜索了Materials Project数据库中近60000种晶体材料，发现了系列类金刚石的高热导率材料，为高导热材料与器件的设计提供了更多选择与可能。（2）基于贝叶斯优化及蒙特卡洛树的纳米热输运结构优化，将贝叶斯优化及蒙特卡洛树算法和传统的热输运计算进行了结合，实现了纳米结构最佳/最差导热特性的智能化结构设计。该设计方法具有通用性和普适性，可推广到热辐射、光子、磁子等各种复杂能量输运设计体系。这些成功案例表明了材料信息学设计与优化效率高，可有效降低新材料的研发周期和成本。

关键词：材料信息学，热功能材料，数据库搜索，结构设计与优化