【作者】 夏伯才；

【导师】 钱翰城；

【作者基本信息】 重庆大学 ， 材料加工工程， 2004， 博士

【摘要】 由材料成分和工艺预测性能，或由预定性能设计材料成分和工艺，实现材料的计算设计，是材料科学与工程中重要的研究课题之一，但目前尚难以从机理出发通过理论计算设计工程材料。一种可行的方法是，建立数据模型，进行经验设计。由于传统的统计建模难以反映材料中存在的复杂交互作用关系，而用软计算实现材料经验数据建模和优化更为有效，目前正成为新的学科交叉点和研究热点。因此，开展材料经验数据建模与优化研究，不但可以丰富和完善材料设计理论和方法，而且具有重要的实用价值。鉴于此，论文旨在形成一个用于材料经验设计的数据软计算建模与优化框架。主要对以下内容开展研究：改进神经网络学习算法，简化模糊推理系统；改进遗传算法和粒群优化算法；开展软计算在工程材料中的应用研究。在建模与优化理论方面，主要创新与突破表现在：提出了基于贝叶斯权重规范化的差异演化训练算法，改善了多层前馈型网络的学习精度和泛化能力；综合采用模糊聚类、误差回逆学习及基于相似分析的规则库约减，提高了高木—关野(Takagi-Sugeno)型模糊推理系统的可解释性；改进了遗传算法编码方式，并用模糊控制策略调整交叉和变异过程，形成了高效的十进制遗传算法和十进制近似遗传算法，提高了寻优速度和计算效率；用模糊控制和差异演化控制粒群优化参数，形成了高效快速的复合粒群算法，同时将捕食优化等新算法用于材料优化；实现了模糊决策的多目标材料参数优化，形成了基于软计算的材料经验设计理论框架。在实际应用方面，取得的成果包括：建立了基于网形图的冲天炉熔炼工艺模型，在预定熔化率和炉温的模糊限制条件下，以提高热效率为目标，实现了工艺参数的模糊多目标优化；通过ZL101A合金成分－力学性能建模分析，对Si、Mg和Fe含量进行优化，在湿砂型T5状态下，使抗拉强度(b和伸长率(5稳定在250±10MPa和3.0±0.5%；通过A319铸造铝合金固溶处理工艺试验，建立了固溶工艺模型，得到了最佳固溶温度和时间参数，在金属型T6状态下，使抗拉强度稳定在340MPa左右且伸长率在5％以上；通过高碳当量灰铁成分－性能建模，进行成分设计，制备出具有良好工艺性能的高碳当量高强度灰铁；通过建立HT250成分、出铁温度、浇注温度与抗拉强度和硬度之间的关系，用多种遗传算法实现了预定强度和硬度下铁水成分与冲天炉熔炼工艺的综合优化。通过综合采用多种软计算中技术，开展经验数据建模与优化研究，改进现有算法，形成了材料经验数据软计算建模与优化框架体系，并成功用于工业合金材料的成分和工艺设计。 <WP=7>研究形成的技术和方法具有一定的通用性和推广价值，可以应用到其它类似工艺过程的建模分析与优化。更多还原

【Abstract】 One important problem in material science and engineering is the computational design of material, that’s to predict properties through theoretical calculations, or optimize compositions and process parameters of material with prescribed properties. However, for most industrial materials, it’s difficult to design compositions and processes from physical principle. So, a practical alternative is empirical design，that is to develop data-driven models directly from industrial data. For traditional statistical analyses could not deal with the complexity in material data, the soft computing has been used in material modeling and optimization recently， and its applications in material are becoming a new trend in multi-discipline researches. Therefore, the researches on data-driven soft computing modeling and optimization could provide new practical methods, and then enrich methodology in material design theory。The goal of this work is to develop a systematic data-driven modeling and optimization framework for empirical design of material, so it focuses on solving the following problems. First, an attempt is made to improve the generability of neural network and the interpretability of fuzzy system. Then, combining empirical models with improved optimization algorithms, a systematic data-driven soft computing modeling and optimization framework is to be developed. And finally, all of the above would be validated in material processing practices based on industrial data.After systematic researches on the data-driven modeling and optimization, thesis’s main theoretical contributions are as following. Firstly, using Bayesian complexity regularization for weight decay in the error back-propagation learning procedure, differential evolution method is used in determination of optimal multilayer perceptrons networks with better approximation and generalization. Secondly, an adaptive neuro-fuzzy system consisted of reduced interpretable Takagi-Sugeno type inference rules is also proposed to improve the predictability and generalization. In fuzzy inference system, the structure identification and parameter optimization are carried out automatically and efficiently by the combined use of fuzzy clustering, adaptive back-propagation learning, and similarity analysis-based model simplification. Thirdly, instead of binary string encoding, more quickly genetic algorithms with improved genetic operators are developed, in which digits vary over the numbers 0, 1, 2, . . . 9 and fuzzy controllers are also used to adapt the crossover and mutation. A hybrid particle <WP=9>swarm optimization (PSO) exhibiting higher success rates is developed. The hybrid PSO utilizes fuzzy control method and differential evolution algorithm to determine the appropriate set of parameters during the optimization. At the same time, bacterial foraging optimization (BFO) is also applied to the material processing. And finally, a systematic modeling and fuzzy multi-objects optimization framework is proposed, in which by the use of fuzzy multi-attributes material design method one could describe uncertain and imprecise information in materials.In practices, integrating adaptive neruo-fuzzy inference models into optimization procedure, a number of industrial problems related to material processing have been solved. With operating rules generated directly from cupola melting experimental data, a set of neuro-fuzzy inference models has been developed to determine the optimal blast and carbon rate for the lowest energy consumption with fuzzy constrained tap temperature and melting rate. After modeling the correlation between mechanical properties and compositions of ZL101A cast aluminum alloy, with the optimized compositions, the ultimate tensile strength and elongation are controlled within the range of 250(10MPa and 3.0(0.5%。Based on solution treatment experiments of A319 cast aluminum alloy, neuro-fuzzy models have been obtained to describe the relationship between mechanical properties and solution parameters. Combined the fuzzy models with optimization pr更多还原