| The computational models of complex structures such as steel plate shear walls(SPSWs)have been widely studied,while most traditional models require trade-offs in terms of accuracy,efficiency and applicability.In view of these deficiencies,this dissertation conducts systematic research based on the computational models of SPSWs with the aim of high fidelity,high efficiency as well as generality.The study starts from the finite element techniques,and furthermore,explores a new structural intelligenet computational framework constituted by state-of-the-art deep learning models.The major research achievements are as follows:(1)A high-fidelity elasto-plastic constitutive model SSTHU was developed for structural steel based on the cyclic coupon tests.A semi-automatic and efficient calibration method was put forward to deal with massive entangled material-dependent parameters.The proposed model served as the accuracy guarantee for the finite element model of SPSWs.(2)A highly-efficient algorithm inverse-P-norm method was proposed to accommodate all spatial implementation of arbitrary J2 elasto-plastic constitutive models,wherein the efficiency was realized by yielding a univariate scalar equation and the explicit analytical expression of consistent elasto-plastic moduli.The algorithm was further equipped with object-oriented programming to enhance the maintainability of the codes and the scalability of the underlying models.The inverse-P-norm method served as the efficiency guarantee for the finite element model of SPSWs.(3)Numerical simulation of numerous SPSWs with all kinds of construction details were conducted by virtue of the developed finite element model.Although the results indicated that the model was basically able to integrate accuracy,efficiency and the generality,it still had a large space for optimization.Therefore,the mathematical mechanism of the traditional computational models were revealed through both the macro-member level and the micro-material level,which demonstrated the feasibility as well as the advantanges of the deep learning technologies.(4)The main characteristics of structural computational task were analyzed and three points were concluded: ultra-long sequence length,siginificant memory effect,and causal autoregression.Considering these three intrinsic features,Transformer architecture was introduced,and the FAVOR+ mechanism,local multi-head attention mechanism,and reversible network technique were adopted to form Mechformer model,which was applicable to any history-dependent mechanical response prediction problem.(5)A uniform data interface was designed for the static features of structures and a deep learning model PADCN was established for dealing with the representation learning of arbitraty static features.In order to alleviate the deficiency of experimental data,a series of algorithms was put forward to augment training samples.Meanwhile,model-agnostic meta-learning was introduced to implement few-shot learning,based on which the Meta-Mechformer framework was proposed.The validation by SPSWs experiments showed that the model was far more accurate and highly-efficient than the traditional models. |
