【作者】 李广召；

【导师】 朱昶胜；

【作者基本信息】 兰州理工大学 ， 物联网工程， 2024， 硕士

【摘要】 机器学习作为有效可靠的计算技术在材料工程学中的应用非常广泛。在合金材料的生产加工中,合金材料的弹性性能和晶格常数等特性在产品的设计以及产品的质量方面发挥着重要的作用。传统的计算与仿真的方法已经无法满足工业上对于合金材料的应用需求,利用机器学习(Machine learning,ML)和材料科学相结合的方法来辅助合金材料弹性模量和晶格常数的模拟和预测具有重要意义。基于此,本文主要研究内容如下:(1)利用320个镁铝铜合金样本为试验材料,采用密度泛函理论(Density Functional Theory,DFT)方法计算弹性模量,对计算的原始数据进行归一化处理,构建一个基于镁铝铜合金包含晶胞体积、晶胞原子个数、晶体密度、晶体维度、形成能、吉布斯自由能同数量级的弹性模量数据集。建立基于布谷鸟算法(Cuckoo Search,CS)优化随机森林(Random Forest,RF)的弹性模量预测模型,利用构建弹性模量数据集训练模型,并对镁铝铜合金的弹性模量进行预测研究。实验结果表明,对于预测合金的弹性模量,CS-RF在提高预测精度的前提下,计算效率也有明显的优势。CS-RF模型相较于传统的随机森林模型的运算速度提高3倍。(2)基于Pymatgen平台,通过Materials Project数据库中的REST API端口采集2667个包含镁基、铝基、铜基面心立方合金样本数据,对试验材料数据进行归一化处理,构建了一种基于面心立方合金包含原子数量、多种范式、轨道电子、离子能、形成能、能带中心等多维结构的晶格常数数据集。建立粒子群算法(Particle Swarm Optimization,PSO)优化的支持向量机模型(Support Vector Machine,SVM)和遗传算法(Genetic Algorithm,GA)优化的支持向量机模型对面心立方合金晶格常数进行预测研究,并进行对比分析。实验结果表明,在预测面心立方合金的晶格常数方面,PSO-SVM在预测精度可以达到99.003%的高精度,计算效率相较于GA-SVM模型具有明显的优势,证明PSO-SVM是一种先进的机器学习方法。(3)利用CS算法和PSO算法的每次迭代寻优结果,获取每次迭代的种群信息,并筛选出效果较好的解作为精英种群,利用多个精英种群进行种群间的交叉、选择、进化等方法构建包含多个极值解的多精英种群。为了解决寻优算法会陷入局部最优解的问题,建立多精英种群算法(Multi-elite Populations,MEP)对CS算法和PSO算法的种群进行优化处理,并对弹性模量和晶格常数进行预测研究,使得预测模型的初始寻优速度提高,提高算法迭代的效率,提高算法预测的精确度。更多还原

【Abstract】 Machine learning as an effective and reliable computing technology,and is currently widely used in materials engineering.In the production and processing of alloy materials,the elastic properties and lattice constants of alloy materials play a very important role in the design of products and the quality of products.Traditional calculation and simulation methods can no longer meet the application needs of alloy materials in the industry.Therefore,it is necessary to use machine learning(ML)methods to simulate and predict the various properties and properties of alloy materials.Therefore,it is of great significance to use the combination of machine learning and materials science to assist in the simulation and prediction of the elastic modulus and lattice constant of alloy materials.Based on this,the main research contents of this paper are as follows:(1)Using 320 samples of magnesium-aluminum-copper alloy as test materials,the elastic modulus was calculated by density functional theory(DFT)method,and the calculated original data were normalized to construct a dataset of elastic modulus based on the same order of magnitude as unit cell volume,unit cell atoms,crystal density,crystal dimension,formation energy,and Gibbs free energy of magnesium-aluminumcopper alloy.An elastic modulus prediction model based on cuckoo algorithm Optimized Random Forest(CS-RF)was established,and the model was trained on the elastic modulus dataset of magnesium-aluminum-copper alloy,and the elastic modulus of magnesium-aluminum-copper alloy was predicted by using the model.Experiments show that CS-RF has obvious advantages in computational efficiency for predicting the elastic modulus of alloys under the premise of improving the prediction accuracy.Compared with the traditional random forest(RF)model,the CS-RF model is 3 times faster.(2)Using the Pymatgen platform to download 2667 samples of magnesium-based,aluminum-based and copper-based face-centered cubic alloys through the REST API port in the Materials Project database,the test material data were normalized to construct a lattice constant dataset based on the multi-dimensional structure of the facecentered cubic alloy including atomic number,multiple paradigms,orbital electrons,ion energy,formation energy,and band center.The support vector machine model optimized by particle swarm optimization(PSO-SVM)and the support vector machine model optimized by genetic algorithm(GA-SVM)were established to predict the lattice constant of face-centered cubic alloys,and the comparative analysis was carried out.Experiments show that PSO-SVM is an advanced machine learning method in predicting the lattice constant of face-centered cubic alloys,and PSO-SVM can achieve a high accuracy of 99.003% in prediction accuracy,and the computational efficiency has obvious advantages compared with the GA-SVM model.(3)Using the optimization results of each iteration of CS algorithm and PSO algorithm,the population information of each iteration was obtained,and the solution with better effect was screened out as the elite population,and the multi-elite population containing multiple extreme solutions was constructed by using multiple elite populations for crossover,selection,and evolution between populations.In order to solve the problem that the optimization algorithm will fall into the local optimal solution,the Multi-Elite Population Algorithm(MEP)is established to optimize the populations of the CS algorithm and PSO algorithm,and the elastic modulus and lattice constant are predicted,so that the initial optimization speed of the prediction model is improved,the efficiency of algorithm iteration is improved,and the accuracy of algorithm prediction is improved.更多还原