题目：人工智能加速的计算和实验材料设计

报告人：刘轶 教授（上海大学 材料基因组工程研究院 Email: yiliu@shu.edu.cn）

时间：2025年6月10日（星期二） 14:30

地点：师昌绪楼408会议室

联系人：材料设计与计算研究部 刘培涛

报告人简介：

Prof. Yi LIU obtained his Ph. D. degree at Materials Science and Engineering at Institute of Metal Research in China in 1997. Then he has worked in the field of computational materials science at Nagoya University,Japan (1997-2002);Juelich Research Center,Germany (2002-2003);University of Western Ontario,Canada (2003-2005);California Institute of Technology,US (2006-2012). He is a professor at Materials Genome Institute and Department of Physics at Shanghai University (2015-present) after working at the School of Materials Science and Engineering,the University of Shanghai for Science and Technology (2012-2015). His current research interests focus on the multi-paradigm materials design for advanced alloys,energy materials,and nanomaterials by combining computation (density functional theory and reactive force field molecular dynamics simulations),AI/machine learning,and high-throughput experiment approaches.

报告摘要：

In this talk I will show how to apply artificial intelligence to accelerate computational and experimental materials design.

Part I

“What you need is pre-attention”:Machine learning with Center-Environment features improves small-data materials design via a pre-defined attention mechanism

The resurgence and widespread application of artificial intelligence generally rely on the combination of big data and deep learning algorithms. However, data in materials science research are often scarce, incomplete, and highly uncertain, posing severe challenges to the search and design within the vast materials parameter space. To enable small-data-driven materials design, we propose a machine learning (ML) method based on a "pre-attention" mechanism. Adhering to the principles of feature engineering, we construct a "Center-Environment" (CE) feature model that reflects core-shell structural characteristics by leveraging domain knowledge in materials science. The CE model introduces the concept of pre-attention by focusing limited data on a feature model with physical significance.

“All you need is attention”. Currently popular Transformer algorithms in large language models require large amounts of data to achieve a multi-head"self-attention"mechanism. In contrast,the CE pre-attention mechanism shifts attention from complex black-box machine learning algorithms to explicit feature models with physical meaning,reducing data requirements while enhancing the transparency and interpretability of machine learning models. We combine CE features with kernel functions or deep machine learning algorithms to construct machine learning models,successfully applying them to studies of bulk materials,surfaces,and local doping systems,involving areas such as new material discovery,surface catalysis,and alloy effects. In this talk,I will mainly introduce the ML-CE modeling of Nb/NbSi alloys for alloying element effects on both stability and mechanical properties.

“What you need is pre-attention”. Comparative studies show that in small-data scenarios,our CE machine learning model exhibits higher accuracy and broader applicability than traditional deep learning models based on graph features. Since CE can be used to describe features of any complex crystal structure,machine learning based on CE features can become an effective and general method for data-driven materials design oriented towards small datasets.

Part II

High-throughput experimental and machine learning optimization of composition and processing for high-strength and high-conductivity copper alloys beyond the scale of thousand samples per year

As the lead frames of electronic chip and electric contact materials,the high-strength and high-conductivity copper alloys need to simultaneously satisfy high mechanical strength and high electrical conductivity, achieving both good mechanical and electrical properties is often challenging due to their inherent contradiction. Furthermore, the comprehensive performance of copper alloys is influenced by a myriad of complex factors such as alloy composition, heat treatment, and rolling deformation processes. Optimizing alloy composition and processing concurrently to meet the multi-objective material performance requirements is a practical necessity for the development and industrial application of new materials. However,due to the vast potential material parameter space,large-scale systematic optimization of composition and processing remains highly challenging.

This work introduces a high-throughput optimization of compositions and processing of multi-component copper (Cu-Zr-Cr) alloys at the scale of thousand samples per year,coupled with machine learning-based performance prediction at the million-level scale. The combined high-throughput experiment and machine learning provide an efficient “composition-processing-performance” holistic optimization capability for the development and industrial application of novel multi-component alloys. By simultaneously tunning key factors such as alloy composition,rolling deformation rate,and aging temperature/time,a total of 1669 copper alloy samples were prepared using a high-throughput arc melting,heat treatment,and rolling system developed in-house,with hardness and electrical conductivity measured for each (3338 experimental data points in total). Machine learning models were constructed based on the high-throughput experimental data to predict hardness and electrical conductivity,with 1159 data points in the training set and 480 in the independent validation set,further extending predictions to a parameter space of one million (1,039,140) material combinations. Finally,copper alloy samples with typical performance were subjected to optical,scanning,and transmission electron microscopy observation to analyze and discuss the relationship between alloy microstructure and performance.