基于U-Net模型的材料动态力学响应预测:织构表征差异的影响OA
Prediction of Dynamic Mechanical Response of Materials Based on U-Net Model:Influence of Texture Representation Differences
利用神经网络进行合金性能的预测以及合金微观结构的反向设计已经成为了解合金性能、开发新型合金的一种新兴手段.织构是合金变形过程中微结构演化的重要影响因素,通常采用无空间关联的离散晶粒取向欧拉角、考虑代表性体积单元(representative volume element,RVE)空间-取向耦合欧拉角、极图/反极图等形式描述.选取何种形式的织构表示方式作为神经网络模型的输入能够发挥模型的最优性能还需进一步验证.为此,以改进后的U-Net模型为主体架构,对比了使用离散欧拉角、空间欧拉角、极图 3 种织构表示方法作为神经网络模型输入对模型性能的影响.采用训练完成的 3 种神经网络模型分别对测试集中的样本进行预测,结果表明,使用极图作为织构表示方法时获得了最优的效果.除此以外,利用训练好的神经网络模型,采用在输出层添加一维卷积的改进方法,预测了合金的宏观应力-应变曲线.相比于传统只使用全连接的方法,该改进显著提高了应力-应变曲线的预测精度.
The utilization of neural networks for the prediction of alloy properties and the inverse design of alloy microstructures has emerged as a novel approach in the industry for understanding material performance and developing new alloys.Texture acts as a critical factor influencing microstructural evolution during alloy deformation.It is typically characterized by spatially uncorrelated discrete grain orientation Euler angles,spatial-orientation coupled Euler angles within a representative volume element(RVE),or pole figures/inverse pole figures.However,identifying which texture representation method serves as the optimal input to maximize the performance of neural network models requires further investigation.Consequently,employing a modified U-Net model as the backbone architecture,this study evaluates and compares the impact of three texture representation methods,including discrete Euler angles,spatial Euler angles,and pole figures,as model inputs on the overall performance of the neural network.The three trained neural network models were individually deployed to predict samples within the test set.The results demonstrate that employing pole figures as the texture representation method yields the optimal performance.Furthermore,the trained neural network models were utilized to predict the macroscopic stress-strain curves of the alloys by incorporating a one-dimensional(1D)convolutional layer at the output stage.Compared to traditional methods relying solely on fully connected layers,this modification significantly enhances the prediction accuracy of the stress-strain curves.
高翔;赵聃;方慧青;王建军;马胜国;王志华
太原理工大学航空航天学院,山西 太原 030024||太原理工大学材料强度与结构冲击山西省重点实验室,山西 太原 030024太原理工大学航空航天学院,山西 太原 030024||太原理工大学材料强度与结构冲击山西省重点实验室,山西 太原 030024太原理工大学航空航天学院,山西 太原 030024||太原理工大学材料强度与结构冲击山西省重点实验室,山西 太原 030024太原理工大学航空航天学院,山西 太原 030024||太原理工大学材料强度与结构冲击山西省重点实验室,山西 太原 030024太原理工大学航空航天学院,山西 太原 030024||太原理工大学材料强度与结构冲击山西省重点实验室,山西 太原 030024太原理工大学航空航天学院,山西 太原 030024||太原理工大学材料强度与结构冲击山西省重点实验室,山西 太原 030024
数理科学
神经网络织构动态应力-应变曲线U-Net模型
neural networktexturedynamic stress-strain curveU-Net model
《高压物理学报》 2026 (10)
52-62,11
国家自然科学基金(12225207,12372364,12172245)
评论