4-8. Machine learning and big data towards ultra-high temperature

Yanhui Zhang*, Alessandro Lunghi, Stefano Sanvito

Trinity College Dublin

Abstract: The severe service environment of high temperature structural materials, such as extremely high temperatures (above 2000 °C), makes experimental measurements very difficult, costly and some time impossible. In this work, we present a tangile simulation tool by integrating the merits of machine learning, first-principles and molecular dynamics, which promises great chances of pushing the limits of high temperature studies. Relying on this research strategy, we discuss the success in developing a highly efficient and accurate SNAP force field for ZrB2. Our machine-learned force field largely outperforms the empirical potentials in terms of predicting thermal, mechanical and transport properties across a broad range of temperatures. The power of this simulation strategy can easily transfer to the study of superalloys and high entropy alloys etc. Our work demonstrates that machine-learning force fields can be used for the simulations of materials in extreme environment where no experimental tools are available. And it highlights a significantly efficient and accurate way on the investigation of engineering materials in the domain of ultra-high temperature, which are significant for many technologies, including aircraft, aerospace, energy and defense industries.

Keywords: Machine learning; Force field; Ultra-high temperature ceramics; Thermal conductivity

机器学习和大数据在超高温材料研究中的应用

张宴会*, Alessandro Lunghi, Stefano Sanvito

Trinity College Dublin

摘要：航空航天和核能工业中的热端部件，如火箭喷嘴和热防护系统，面临着苛刻工况条件的考验。这通常包括超高的温度(火焰温度2700-3500 ℃)、极强的热烧蚀（热流强度1-15 MWm^-2）及化学腐蚀性环境（ H₂O, O, HCl 等），是多场耦合下的复杂现象。受限于表征技术的精度、样品制备质量及高温检测的困难，传统的实验技术很难进行实时和高精度地测测分析。计算模拟研究需在不同尺度上展示材料对热、力场的综合响应。第一性原理(DFT) 、分子动力学(MD)、粗晶(DPD)和有限元(FEM)等所适用的时间和空间尺度不同。值得注意地是，如何实现不同模拟计算方法之间有效的信息交流，对解析材料结构-性能的本征关系起纽带作用。最新的国际研究热点是利用机器学习方法来进行不同尺度的计算模拟方法之间的数据和信息传递。

关键词：机器学习；DFT数据库；高通量计算；分子动力学模拟；超高温