具有历史依赖效应的材料及结构响应预测通用深度学习MechPerformer模型

MechPerformer: a general deep learning model for history-dependent response prediction in structural engineering

针对结构计算分析领域具有历史依赖效应的力学响应预测问题,基于深度学习技术提出了一个通用模型,即MechPerformer模型。相较于自然语言处理领域,结构计算领域具有序列长度极长、记忆效应显著、因果自回归性等特性。围绕其特性,引入Transformer架构搭建了MechPerformer模型,采用了基于随机正交正特征映射的注意力机制、反向网络技术、门控循环单元等以适应结构计算分析应用场景。为验证深度学习模型,利用历史路径强相关的低屈服点钢精细弹塑性本构模型生成了随机循环应力-应变响应数据以开展数值试验,并在其中采用了适用于各种力学量的统一归一化方法——非线性参考值放缩法。试验结果表明,MechPerformer模型能够高精度复现具有强历史依赖效应的非线性响应曲线,且呈现远超理论模型的计算效率,证明了模型的有效性。同时,根据经典弹塑性力学理论,给出了模型的力学意义阐释,增强了MechPerformer模型的可解释性。

To address the history-dependent response prediction problems in structural analysis, a general deep learning model MechPerformer was proposed. Main characteristics of structural computational analysis compared with natural language processing were concluded, i.e., ultra-long sequence length, remarkable memory effect, and causal autoregression. To address these three characteristics, a general deep learning model named MechPerformer based on Transformer architecture was developed, integrating FAVOR+(fast attention via positive orthogonal random features) mechanism, reversible network technique, and gated recurrent unit. In order to validate the proposed model, a numerical experiment based on the random cyclic stress-strain response of a low-yield-point steel constitutive model was carried out, in which a novel data normalization method, named nonlinear reference-value scaling-tailored for arbitrary mechanical variables was proposed. The experimental results verified the effectiveness and the high efficiency of the new model, which successfully reproduced the highly nonlinear history-dependent curves. Furthermore, the feasibility of the combination of the MechPerformer model and finite element technique was discussed. Finally, a qualitative mechanical explanation of the MechPerformer model was provided to enhance the interpretability.