報告題目：Machine learning approaches to materials failure: From the atomic to the macroscale

報 告 人：Michael Zaiser  博導、教授

報告時間：5月22日（星期四）15:00-16:30

報告地點：明德樓B308

報告人簡介：

畢業于德國斯圖加特大學，獲理論物理博士學位。歷任英國愛丁堡大學講師、教授，2012年起任德國埃爾朗根-紐倫堡大學材料系講席教授（Chair Professor）。四川省“天府學者”，教育部“海外名師”。主要研究領域為微納米材料力學及高性能材料，應用統計物理、材料科學、固體力學等理論，研究材料的微結構和缺陷的無序性和隨機性，及對其材料宏觀力學性能的影響。在Science，Nature Physics，Nature Communications，Advanced Energy Materials，Physical Review Letters等國際頂尖期刊上發表學術論文200余篇，累計引用達8000余次，出版專著2部，參與編寫專著5部。

報告內容摘要：

Materials failure is a complex spatio-temporal phenomenon involving processes that range from the interactions of single atoms and the breaking of atomic bonds to stress re-distribution and rupture on macroscopic and even geophysical scales. Machine learning approaches can help us to understand failure processes based on simulation and monitoring data, and to use this information in order to predict when, where and how materials fail. We illustrate this for two examples: (i) On the atomic scale, we show how a combination of classification and image recognition tools can be used to identify atomic environments that are prone to crack nucleation under load, to predict the critical loads for cracks to emerge, and to forecast the ensuing crack paths; (ii) on the scale of macroscopic creep samples of metallic glasses, we show how regression tools can be developed that allow to predict, based on monitoring records, sample specific failure times, and we discuss the importance of establishing and exploiting spatial signatures of incipient failure to obtain accurate forecasts. Based on these examples, we discuss general issues related to feature selection and the importance of incorporating domain knowledge into machine learning approaches.

主辦單位：新能源與材料學院/光伏新能源現代產業學院

油氣藏地質及開發工程全國重點實驗室

科學技術發展研究院

成都新材料學會

氫能制取與高效利用重點實驗室

成都市科技青年聯合會材料能源專委會

能量轉換與儲存先進材料四川省國際科技合作基地

四川省頁巖氣高效開采先進材料制備技術工程研究中心

四川省玄武巖纖維復合材料開發及應用工程技術研究中心