人工神经网络在材料科学中的研究进展

Research Progress of Artificial Neural Networks in Material Science

传统的“试错”型材料研究方法存在周期长、成本高和偶然性大等不足,已经不能满足现代材料研发的需求,提高研发针对性、缩短材料研发周期、降低材料研发成本成为全世界的研究热点。随着数据量的不断累积以及计算机技术的不断发展,数据密集型科学逐渐成为科学研究的第四范式。从大量数据中寻找能反映材料本征的“基因”,是材料现行的研究趋势。人工神经网络方法因具备自学习、联想存储以及高速寻找优化解的能力的优点而被广泛应用于材料科学领域。研究者利用人工神经网络等机器学习模型对材料的试验或理论计算数据进行挖掘,在专家经验和理论指导下转化为可靠的知识并能够辅助智能决策,从而建立材料从微观结构到宏观性能间的一一映射关系。早期,人工神经网络主要被用于寻求材料的宏观参数与材料宏观性能之间的关系,如材料的成分设计,加工过程的工艺参数优化,以及寻找影响材料使用性能的环境参数;人工神经网络通过对第一性原理计算结果进行学习,被用于描述原子尺度下体系之间的作用关系,以此实现计算速度与精度的平衡;而卷积神经网络等深度神经网络方法在图像处理上的独到优势,使得其在材料表征领域得到了更广泛的应用,如SEM、TEM中微结构识别与重构。借助人工神经网络等方法,实现材料微观、介观到宏观性能之间跨尺度的联系,是实现材料设计这一终极目标的可能途径。本文回顾了人工神经网络的发展历史,对目前材料领域应用最为广泛的BP神经网络与卷积神经网络原理进行了阐释,综述了人工神经网络在材料宏观性能、计算模拟、材料表征等领域的应用,探讨了人工神经网络在材料领域应用存在的不足,最后对未来的发展趋势进行了展望。

The traditional "trial and error" material research methods have the disadvantages of long cycle,high cost and great contingency,which can’t meet the needs of modern material research and development. It has become a research hotspot all over the world to improve the pertinence of research and development,shorten the cycle of material research and development,and reduce the cost of material research and development. With the continuous accumulation of data and the continuous development of computer technology,data intensive material scientific discovery has gradually become the fourth paradigm of scientific research. It is the current research trend of materials to find the"gene"that can reflect the material’s intrinsic characteristics from a large number of data. The artificial neural network method is widely used in the field of materials science because of its advantages of self-learning,associative storage and high-speed search for optimal solution. Researchers use machine learning models such as artificial neural network to mine the experimental or theoretical calculation data of materials under the guidance of expert experience and theory,it can be transformed into reliable knowledge and can assist intelligent decision-making,so as to establish a one-to-one mapping relationship between microstructure and macro performance of materials.In the early days,artificial neural network was mainly used to find the relationship between the macro parameters of materials and the macro performance of materials,such as the composition design of materials,the optimization of process parameters,and the search of environmental parameters that affect the performance of materials. With the development of computer technology and the rise of computer simulation,artificial neural network was used to learn the calculation results of the first principle. It is used to describe the interaction between the systems at the atomic scale,so as to achieve the balance between the calculation speed and accuracy. The convolution neural network methods,such as a deep neural network method,has unique advantages in image processing,which makes it more widely used in the field of material characterization,such as microstructure identification and reconstruction in SEM and TEM. With the help of artificial neural network and other methods,it is a possible way to realize the ultimate goal of material design to realize the cross scale relationship between micro,meso and macro properties of materials.This paper reviews the development history of artificial neural network,explains the principle of BP neural network and convolution neural network which are widely used in the field of materials at present,summarizes the application of artificial neural network in the field of material macro performance,calculation simulation,material characterization,etc.,probes into the shortcomings of the application of artificial neural network in the field of materials,and finally discusses the development trend of artificial neural network in materials research.