【作者】 姜雪；

【导师】 尹海清；

【作者基本信息】 北京科技大学 ， 材料科学与工程， 2020， 博士

【摘要】 科学研究第四范式的提出,宣告着数据密集型的科学研究方法时代的到来。机器学习和人工智能技术的发展,为日趋复杂的材料设计提供了崭新的途径。数据库与机器学习技术,能够揭示设计过程与材料宏观性能的关系,实现实验过程优化,改善依靠科学家直觉和大量的“尝试法”的材料研发思路,形成数据驱动的研究方法,从而加速材料设计与研发。本课题基于数据库和机器学习技术,针对合金材料的组织和性能开展数据驱动的设计方法研究。通过分析材料实验表征数据、计算模拟数据和工业生产数据的特征,研究材料科学数据库构建方法;同时分别针对实验表征数据、计算模拟数据和工业生产数据,开展相关性分析、特征选择、分类和回归分析、集成学习和深度学习等数据挖掘和机器学习技术在优化合金设计方法上的应用研究。基于公开发表的实验数据研究并建立了高温合金晶格错配度模型,基于海量计算模拟数据研究并建立三元合金准相平衡过程的定量设计模型,实现了数据驱动的合金组织参数的定量预测新方法;基于工业数据的高维小样本特点,探索了机器学习与多尺度计算相结合的数据降维新方法,成功用于钢铁工业生产环境下的力学性能高精度预测,为金属结构材料的设计提供了新的方法和思路。具体如下:(1)通过分析材料科学数据的特性,建立了满足材料实验表征数据、计算模拟数据和工业生产数据管理需求的数据定义方法、知识的定义和表示方法,为材料科学数据的分析和利用提供基础支撑,也为基于机器学习的材料设计方法提供了海量异构数据的数据库构建和管理技术。(2)基于公开发表的实验表征数据,研究了镍基单晶高温合金γ’和γ相晶格错配度的高效预测方法,基于支持向量机、序列最小优化、多层感知器等数据挖掘算法,建立化学成分、温度等错配度敏感特征与晶格错配度的关系模型,并从实验与机理模型两方面对数据挖掘模型进行有效性检验,实现镍基单晶高温合金γ’和γ相晶格错配度的准确快速预测。结果表明,多层感知器模型相关系数高,误差低,和传统经验公式相比,预测准确性和效率更有优势。因此,机器学习辅助的晶格错配度预测方法可以在很少的实验和测量的情况下高精度地预测晶格错配度,减少材料设计时间和成本。(3)基于计算模拟数据,研究了多元合金相场模型中的准平衡成分的快速预测方法。以Al-Cu-Mg合金的等温凝固过程为例,基于求解准相平衡方程产生的海量训练样本,利用人工神经网络方法,建立析出相成分、液相成分和相场参数与析出相和液相的准平衡成分的准确定量关系模型。时间开销上,与用最小二乘法求解准相平衡方程相比,机器学习模型只需要1/1000的计算时间。因此,机器学习实现了多元合金相场模拟过程中准相平衡的快速计算,使得相场模拟过程中吉布斯自由能密度耦合热力学模型成为可能。(4)基于工业生产数据,提出了机器学习与多尺度计算相结合的策略,实现“高维小样本”工业环境下的珠光体帘线钢的抗拉强度的准确预测。通过耦合晶粒生长,动态再结晶,温度场和冷却相变计算,将工艺参数空间映射到微观结构空间,实现数据降维目的。利用梯度树提升和高斯过程算法建立先共析铁素体含量,珠光体含量,珠光体层状间距和主要化学成分与抗拉强度的定量关系模型。该模型的平均相对误差小于0.7%,最大相对误差小于2.0%,相比于单纯数据驱动的降维策略可显着提高抗拉强度的预测精度。这种基于机器学习方法的抗拉强度预测方法可以显著加快珠光体钢性能的准确预测,进而加速材料设计和工艺优化。同时,钢铁生产过程中抗拉强度高精度实时预测,为在线质量评估提供了可靠的判断依据,促进企业降低成本和提高工业生产效率。更多还原

【Abstract】 The fourth paradigm of scientific research declares the advent of the era of data-intensive scientific research methods.The development of machine learning and artificial intelligence technology has provided a new way for increasingly complicated material design.Database and machine learning technologies can reveal the relationship between the design process and the macroscopic properties of materials.It can also optimize the experimental process and accelerate material design,instead of relying on the scientist intuition and a large number of "trial and error" experiences.Based on database and machine learning technology,the data-driven prediction methods for the structure and property of alloy materials are developed.By analyzing the characteristics of material experimental data,simulation data and industrial production data,we explore the construction’ method of material scientific database;meanwhile,we carry out data mining and machine learning algorithms,such as correlation analysis,feature selection,classification,regression,ensemble learning and deep learning,to optimize the alloy design methods on the basis of experimental data,calculation simulation data and industrial production data,respectively.The details are as follows.1.By analyzing the characteristics of material science data,data definition methods,knowledge definition and representation methods,and data exchange mechanisms are established for material experimental characterization data,computational simulation data,and industrial production data.These methods provide basic support for the analysis and utilization of materials science data,and also serve as database construction and management techniques for massive heterogeneous data in machine learning.2.Based on the experimental data,we provide a machine learning approach to predict misfit using relevant material descriptors including the chemical composition,dendrite information and measurement temperature and so on.We perform support vector regression,sequential minimal optimization regression and multilayer perceptron algorithms with linear and poly kernels on experimental dataset for appropriate model selecting,and multilayer perceptron model works well for its distinguished prediction performance with high correlation coefficient and low error values.The approach is validated by comparing the predicted lattice misfit with a widely used empirical formula and experimental observation with respect to prediction accuracy.3.Based on the simulation data,we develop a fast and accurate method to predict quasi phase equilibrium via machine learning.Taking the isothermal solidification process of Al-Cu-Mg alloy as an example,the artificial neural network method is used to establish the accurate quantitative relationship model for quasi-equilibrium components of the precipitated phase and liquid phase.The machine learning model only takes 1/1000 of the calculation time comparing with solving the quasi-phase equilibrium equation by the least square method.Its high accuracy and fast speed demonstrate that the neural network model can obtain the quasi phase equilibrium data conveniently in phase field model for multicomponent alloys.4.Based on the industrial data,an effective strategy combining machine learning with multiscale calculation is promoted to construct tensile strength model for pearlitic steel wires.We transform the feature space by coupling of grain growth,dynamic recrystallization,temperature field and cooling phase transition calculations,mapping process space to microscopic structure space.The proeutectoid ferrite content,pearlite content,pearlite lamellar spacing and main composition are used for modeling by Gradient Tree Boosting and Gaussian Process algorithm.It significantly achieves the excellent performance of accuracy,with the mean relative errors less than 0.7%and the maximum relative errors less than 2.0%.This data-driven,globally-optimized and smart prediction method of tensile strength has advantages over traditional costly and time-consuming experimental trial,and can dramatically accelerate material design and process optimization of pearlitic steel,may also being applicable to other structural materials.The high accuracy of prediction ensures a reliable real-time forecast of tensile strength and provides a judgment basis for online quality assessment,which makes great contribution to lowering costs and improving efficiency of industrial production.更多还原