THE BUCKLING ANALYSIS OF THIN-WALLED STRUCTURES BASED ON PHYSICS-INFORMED NEURAL NETWORKS
-
摘要:基于物理信息神经网络(PINN)建立了一种求解薄壁结构屈曲非线性控制方程组的方法. 薄壁结构的非线性控制方程可由挠度和应力函数表示成复杂的4阶非线性偏微分方程组, 使用物理信息神经网络(PINN)解法可以克服传统数值方法对求解域网格的依赖性. 文中建立的神经网络模型根据基于加权的均方误差的损失函数更新网络参数, 并用弧长法迭代的思想进行外层迭代控制以应对屈曲问题的迭代特性. 将弧长法, 硬边界条件, 基于预训练的权重调整策略, 以及自适应激活函数策略融合进网络优化的过程中使得PINN能够更为高效地求解线性与非线性屈曲问题. 文章对两种典型的薄壁结构进行了屈曲模态和带有缺陷的非线性后屈曲问题求解, 并将神经网络获得的解和有限元结果进行了对比. 结果分析表明, 物理信息神经网络方法能够在不需要标签数据的前提下对薄壁结构的屈曲问题进行有效分析, 并且给予的额外标签数据能够提高此方法的求解效率. 该方法虽较成熟的有限元解法收敛速度较慢, 但不需要对求解域进行人为的前处理, 有一定工程应用可行性.Abstract:This paper proposes a method based on the physics-informed neural networks (PINN) for solving the thin-walled structure buckling problem. The governing equations of thin-walled buckling can be expressed as a fourth-order nonlinear partial differential equation system with the in-plane displacement and stress functions as variables. The PINN solution can overcome the dependence of traditional numerical methods on the mesh partition of the computational domain and conduct mesh-free calculations on the entire computational domain. The neural network model presented in the paper utilizes a weighted mean square error loss function composition for updating network parameters and employs the arc-length method for the outer-loop iteration control to deal with the iteration characteristic of buckling problems. The incorporation of the arc-length method, hard constraints, weight adjustment strategy based on trial-and-error pre-training, and self-adaptive activation function strategy enables PINN to solve linear and nonlinear buckling problems effectively. Two types of problems are investigated in the study, including buckling mode analysis and nonlinear post-buckling problems with geometry deficiencies. A comparison is made between the solutions obtained from the neural network and finite element results. The results demonstrate the efficacy of the proposed method in accurately solving both linear and nonlinear buckling problems in thin-walled structures, highlighting its potential applications in structural engineering and design optimization. The research results show that the physics-informed neural network can effectively analyze the buckling problem of thin-walled structures without requiring artificial preprocessing of the computational domain. Additionally, PINN retains the traditional characteristic of normal DNNs and can accept labeled data for faster calculations. The paper shows that the labeled buckling mode data can accelerate the convergence of the network. The drawback of PINN is that it converges slower than the mature finite element method, but the feature of requiring no artificial preprocessing of the solution domain before the training process makes PINN feasible in engineering.
-
算法1: 基于预训练的损失函数权重的PINN训练过程 Input: 预训练任务最大迭代数$ite{r_{{\rm{pre}}}}$, 主要训练任务最大迭代数$ ite{r_{\max }} $, 数量级超参$ {\boldsymbol{\alpha }} $ 1: 初始化神经网络权重和偏置向量$\left[{{\boldsymbol{W}}}^{i},{{\boldsymbol{b}}}^{i}\right], i\in La$ 2: ${\left( {\tilde W,\tilde F} \right)_{ite{r_{{\rm{pre}}}}}} = \Im (x,y,\varTheta )$(神经网络预训练${N_{{\rm{pre}}}}$次) 3: $\begin{gathered}Los{s_{ite{r_{{\rm{pre}}}}}} = {\left\| {{\boldsymbol{MS}}{{\boldsymbol{E}}_P}} \right\|_1} + {\left\| {{\boldsymbol{MS}}{{\boldsymbol{E}}_B}} \right\|_1} + {\left\| {{\boldsymbol{MS}}{{\boldsymbol{E}}_I}} \right\|_1} \\ \lg \left( {{{\boldsymbol{w}}_P}} \right) = \left[ {\boldsymbol{\alpha }} \right] - \left[ {{\text{lg}}{{\left( {{\boldsymbol{MS}}{{\boldsymbol{E}}_P}} \right)}_{{N_{{\rm{pre}}}}}}} \right] \\ \lg \left( {{{\boldsymbol{w}}_B}} \right) = \left[ {\boldsymbol{\alpha }} \right] - \left[ {{\text{lg}}{{\left( {{\boldsymbol{MS}}{{\boldsymbol{E}}_B}} \right)}_{{N_{{\rm{pre}}}}}}} \right] \\ \lg \left( {{{\boldsymbol{w}}_I}} \right) = \left[ {\boldsymbol{\alpha }} \right] - \left[ {{\text{lg}}{{\left( {{\boldsymbol{MS}}{{\boldsymbol{E}}_I}} \right)}_{{N_{{\rm{pre}}}}}}} \right] \\ {\text{ }} \\ \end{gathered}$ 4: for $ epoch = 0 $ to $ epoch = ite{r_{\max }} $ do:(主训练任务) { ${\left( {\tilde W,\tilde F} \right)_{{\rm{epoch}}}} = \Im (x,y,\varTheta )$(神经网络前向传播) $Los{s_{{\rm{epoch}}}} = {{\boldsymbol{w}}_{{P}}}{\boldsymbol{MS}}{{\boldsymbol{E}}_P} + {{\boldsymbol{w}}_{{B}}}{\boldsymbol{MS}}{{\boldsymbol{E}}_B} + {{\boldsymbol{w}}_{{I}}}{\boldsymbol{MS}}{{\boldsymbol{E}}_I}$ Update $ \varTheta $(神经网络反向传播) } End for 算法2: 基于弧长法的PINN近似解增量求解算法 Input: 利用抽样得到求解板壳上的点集$ \left( {{x_i},{y_i}} \right) $, 初始载荷系数$ LP{F_0} $, 误差容限$ \varepsilon $, 迭代次数上限$ ite{r_{\max }} $ Output: 屈曲偏微分控制方程近似挠度$ \tilde W\left( {x,y} \right) $及其最大值输出$ {W_{\max }} $, 对应的载荷因子$ LP{F_{\max }} $ 1: 构造一个前馈神经网络 $ \tilde W = \Im (x,y,LPF,\varTheta ) $, 其拥有$ La $个隐藏层, 每层有$ {N_u} $个神经元, 训练批次数为$ ite{r_{\max }} $ 2: 进行PDE求解域内和求解边界上的撒点, 域内为$ {N_P} $个, 域边界上为$ {N_B} $个, 采样点总数为$ {N_S} = {N_P} + {N_B} $个 3: $ Loss = {{\boldsymbol{w}}_P}{\boldsymbol{MS}}{{\boldsymbol{E}}_P} + {{\boldsymbol{w}}_B}{\boldsymbol{MS}}{{\boldsymbol{E}}_B} + {{\boldsymbol{w}}_I}{\boldsymbol{MS}}{{\boldsymbol{E}}_I} $ 4: $ \tilde {{W_0}}\left( {{x_i},{y_i}} \right) = \hat u\left( {{x_i},{y_i},LP{F_0},\varTheta } \right) $ 5: while $ \left( {\left| {LP{F_j} - LP{F_{j - 1}} > \varepsilon } \right|} \right)\& j \leqslant ite{r_{\max }} $ do: { $\Delta LP{F_{j + 1}} = \dfrac{{ - \displaystyle\sum\limits_{s = 1}^{{N_S}} {\left[ {\tilde {{W_j}}\left( {LP{F_j},{x_s},{y_s}} \right)\Delta \tilde {{W_j}}\left( {LP{F_{j + 1}},{x_s},{y_s}} \right)} \right]} }}{{{\omega ^2}LP{F_j} - \displaystyle\sum\limits_{s = 1}^{{N_S}} {\left[ {\tilde {{W_j}}\left( {LP{F_j},{x_s},{y_s}} \right)} \right]\frac{{\partial F}}{{\partial w}}\Delta LP{F_j}} }}$ $ LP{F_{j + 1}} = LP{F_j} + \Delta LP{F_{j + 1}} $ $ {\tilde W_{j + 1}}\left( {{x_i},{y_i}} \right) = \Im \left( {{x_i},{y_i},LP{F_{j + 1}},\varTheta } \right) $ } 6: return最佳网络参数 $ {\varTheta ^*} $ 表 1第一阶屈曲模态对应的外载对比
Table 1.The buckling load of the first buckling mode
Boundary conditions PINN results/
(N·m−1)FEM results/
(N·m−1)Relative error SSSS 1466.2 1537.6 4.86% SCSC 4092.1 4048.5 1.07% SCSF 893.2 868.7 2.82% 表 2带缺陷的圆筒壳屈曲行为最大比例载荷因子$ \left( {LP{F_{\max }}} \right) $与无量纲挠度$ \left( {{w_{\max }}} \right) $误差对比
Table 2.The $ LP{F_{\max }} $and $ {w_{\max }} $of shell buckling with initial deficiency and their relative error
Network parameters $ LP{F_{\max }} $ $ {w_{\max }} $ Relative
error$ \left( {LP{F_{\max }}} \right) $Relative
error$ \left( {{w_{\max }}} \right) $L= 3,Nu= 40,
Adam0.522 3.3 5.66% 4.76% L= 3,Nu= 40,
L-BFGS0.455 3.64 8.57% 15.5% L= 4,Nu= 40,
Adam0.480 2.97 2.92% 6.06% L= 4,Nu= 40,
L-BFGS0.432 2.88 14.35% 9.38% L= 4,Nu= 50,
Adam0.531 2.77 7.48% 13.7% L= 4,Nu= 50,
L-BFGS0.399 2.99 23.8% 5.35% 表 3拥有标签数据训练得到的第一阶屈曲载荷
Table 3.The buckling load of the first buckling mode trained with labeled data
Boundary conditions PINN results/(N·m−1) FEM results/×103% Relative error/% SSSS 1721.8 1583.1 8.76 SCSC 4388.2 4169.4 5.24 SCSF 1009.5 894.7 12.77 Variable name Definition $ a $ shell structure dimension $ h $ shell structure thickness $ E $ modulus of elasticity $ \nu $ Poisson's ratio $ P $ shell edge load $ U,V,W $ in-plane displacement $ \tilde W $ fit displacement generated by NN $ Q $ middle surface stress $ F $ airy stress function $ D $ bending stiffness $ {\varepsilon ^0} $ middle surface strain $ La $ NN hidden layer number $ {N_u} $ neuron number for hidden layer $ {\boldsymbol{W}} $ weight matrix for NN $ {\boldsymbol{b}} $ bias for NN $ {\sigma _i} $ activation function for NN $ {\boldsymbol{MSE}} $ mean square error vector $ \varTheta $ hyperparameters for NN ${ {\boldsymbol{w} }_{{P} } }{{,\;} }{ {\boldsymbol{w} }_{{B} } }{{,\;} }{ {\boldsymbol{w} }_{{I} } }$ weights vector for loss function $ {N_P} $ sample points inside the PINN domain $ {N_B} $ sample points on the PINN domain boundary $ {N_S} $ total sample points of PINN $ {N_T} $ test points of PINN $ LPF $ load proportional factor $ lr $ PINN learning rate $ite{r_{{\rm{pre}}} }$ maximum iteration for PINN pre-training $ ite{r_{\max }} $ maximum iteration for PINN training $ {a_{mn}} $ deficiency coefficient $ R $ cylindrical shell radius $ L $ cylindrical shell height $ {\bar Z _B} $ batdorf parameter Numerical example $ \alpha $ ${ {\boldsymbol{w} }_{{P} } }$ ${ {\boldsymbol{w} }_{{B} } }$ plate, SSSS −4 [1,1] [104,104,104,104] plate, SCSC −4 [1,1] [104,103,104,103] plate, SCSF −4 [1,1] [104,103,104,102] shell −4 [104, 10−2] [104,102,102,102,105] Hyperparameter setup Number of random seeds Variance of lg (Loss) $L = 3, {N}_{u} = 60\;\; {\rm{Adam}}\;\; \left[{w}_{P},{w}_{B}\right] = \left[1,1\right]$ 10 1.900 6 $L = 3, {N}_{u} = 60\;\; {\rm{Adam}}\;\; \left[{w}_{P},{w}_{B}\right] = {\left[{w}_{P},{w}_{B}\right]}_{{\rm{epoch}}}$ 10 0.880 6 $L = 3, {N}_{u} = 40\;\;{\rm{Adam}}$ 10 0.931 0 $L = 3, {N}_{u} = 50\;\; {\rm{Adam}}$ 10 0.726 7 $L = 3, {N}_{u} = 70\;\; {\rm{Adam}}$ 10 0.372 5 $L = 3, {N}_{u} = 80\;\; {\rm{Adam}}$ 10 0.883 0 $L = 4, {N}_{u} = 60\;\; {\rm{Adam}}$ 10 0.590 7 $L = 5, {N}_{u} = 60\;\; {\rm{Adam}}$ 10 0.974 5 $L = 3, {N}_{u} = 60\;\; {\rm{Adagrad}}$ 10 0.933 3 $L = 3, {N}_{u} = 60\;\;{\rm{Adadelta}}$ 10 0.794 1 $L = 3, {N}_{u} = 60\;\; {\rm{rmsprop}}$ 10 0.490 9 $L = 3, {N}_{u} = 60\;\; {\rm{L-BFGS}}$ 10 1.041 9 $L = 3, {N}_{u} = 60\;\; {\rm{LAAF} } = \text{relu}$ 10 1.451 8 $L = 3, {N}_{u} = 60\;\; {\rm{LAAF}} = \mathrm{tanh}$ 10 0.652 9 -
[1] 查文舒, 李道伦, 沈路航等. 基于神经网络的偏微分方程求解方法研究综述. 力学学报, 2022, 54(3): 543-556 (Zha Wenshu, Li Daolun, Shen Luhang, et al. Review of neural network-based methods for solving partial differential equations.Chinese Journal of Theoretical and Applied Mechanics, 2022, 54(3): 543-556 (in Chinese)Zha Wenshu, Li Daolun, Shen Luhang, et al. Review of neural network-based methods for solving partial differential equations.Chinese Journal of Theoretical and Applied Mechanics, 2022, 54(3): 543-556 (in Chinese) [2] 陈苏, 丁毅, 孙浩等. 物理驱动深度学习波动数值模拟方法及应用. 力学学报, 2023, 55(1): 272-282 (Chen Su, Ding Yi, Sun Hao, et al. Methods and applications of physical information deep learning in wave numerical simulation.Chinese Journal of Theoretical and Applied Mechanics, 2023, 55(1): 272-282 (in Chinese)Chen Su, Ding Yi, Sun Hao, Zhao Mi, et al. Methods and applications of physical information deep learning in wave numerical simulation.Chinese Journal of Theoretical and Applied Mechanics, 2023, 55(1): 272-282 (in Chinese) [3] Margenberg N, Hartmann D, Lessig C, et al. A neural network multigrid solver for the Navier-Stokes equations.Journal of Computational Physics, 2022, 460: 110983doi:10.1016/j.jcp.2022.110983 [4] Shin J, Park SK, Kim J. A hybrid FEM for solving the Allen–Cahn equation.Applied Mathematics and Computation, 2014, 1(244): 606-612 [5] Xie J, Li M. A fast BDF2 Galerkin finite element method for the one-dimensional time-dependent Schrödinger equation with artificial boundary conditions.Applied Numerical Mathematics, 2023, 168: 89-106 [6] Cybenko G. Approximation by superpositions of a sigmoidal function.Journal Mathematics of Control,Signals,and Systems, 1989, 2(4): 303-314doi:10.1007/BF02551274 [7] Hornik K. Approximation capabilities of multilayer feedforward networks.Neural Networks, 1991, 4(2): 251-257doi:10.1016/0893-6080(91)90009-T [8] Margossian CC. A review of automatic differentiation and its efficient implementation.Wiley Interdisciplinary Reviews:Data Mining and Knowledge Discovery, 2019, 9(4): e1305doi:10.1002/widm.1305 [9] Raissi M, Perdikaris P, Karniadakis GE. Physics informed deep learning (part I): Data-driven solutions of nonlinear partial differential equations.arXiv, 2017, 1711: 10561 [10] Raissi M, Perdikaris P, Karniadakis GE. Physics informed deep learning (part II): Data-driven discovery of nonlinear partial differential equations.arXiv, 2017, 1711: 10566 [11] Raissi M, Perdikaris P, Karniadakis GE. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations.Journal of Computational Physics, 2019, 378: 686-707doi:10.1016/j.jcp.2018.10.045 [12] Baydin AG, Pearlmutter BA, Radul AA, et al. Automatic differentiation in machine learning: a survey.Journal of Machine Learning Research, 2018, 18(153): 1-43 [13] 李野, 陈松灿. 基于物理信息的神经网络: 最新进展与展望. 计算机科学, 2022, 49(4): 254-262 (Li Ye, Chen Songchan. Physics-informed neural networks: recent advances and prospects.Computer Science, 2022, 49(4): 254-262 (in Chinese)Ye L, Songcan C. Physics-informed neural networks: recent advances and prospects.Computer Science, 2022, 49(4): 254-262 (in Chinese) [14] Grohs P, Hornung F, Jentzen A, et al. A proof that artificial neural networks overcome the curse of dimensionality in the numerical approximation of black-scholes partial differential equations. American Mathematical Society, 2023 [15] 陈新海, 刘杰, 万仟等. 一种改进的基于深度神经网络的偏微分方程求解方法. 计算机工程与科学, 2022, 44(11): 1932-1940Chen Xinhai, Liu Jie, Wan Qian, et al. An improved method for solving partial differential equations using deep neural networks.Computer Engineering & Science,2022, 44(11): 1932-1940 (in Chinese) [16] Fang Z. A high-efficient hybrid physics-informed neural networks based on convolutional neural network.IEEE Transactions on Neural Networks and Learning Systems, 2022, 33(10): 5514-5526doi:10.1109/TNNLS.2021.3070878 [17] Pang G, Lu L, Karniadakis GE. fPINNs: fractional physics-informed neural networks.SIAM Journal on Scientific Computing, 2019, 41(4): A2603-A2626doi:10.1137/18M1229845 [18] Xiang Z, Peng W, Zheng X, et al. Self-adaptive loss balanced physics-informed neural networks for the incompressible Navier-Stokes equations.Neurocomputing, 2022, 496: 11-34doi:10.1016/j.neucom.2022.05.015 [19] Shukla K, Jagtap AD, Karniadakis GE. Parallel physics-informed neural networks via domain decomposition.Journal of Computational Physics, 2021, 447(0021-9991): 110683 [20] Tao F, Liu X, Du H, et al. Physics-informed artificial neural network approach for axial compression buckling analysis of thin-walled cylinder//AIAA Scitech 2020 Forum. Orlando, FL. American Institute of Aeronautics and Astronautics, 2020 [21] Li W, Bazant MZ, Zhu J, et al. A physics-guided neural network framework for elastic plates: Comparison of governing equations-based and energy-based approaches.Computer Methods in Applied Mechanics and Engineering, 2021, 383: 113933 [22] Haghighat E, Raissi M, Moure A, et al. A physics-informed deep learning framework for inversion and surrogate modeling in solid mechanics.Computer Methods in Applied Mechanics and Engineering, 2021, 379: 113741doi:10.1016/j.cma.2021.113741 [23] Bastek JH, Kochmann DM. Physics-informed neural networks for shell structures.European Journal of Mechanics-A/Solids, 2023, 97: 104849doi:10.1016/j.euromechsol.2022.104849 [24] Yan CA, Vescovini R, Dozio L. A framework based on physics-informed neural networks and extreme learning for the analysis of composite structures.Computers&Structures, 2022, 265: 106761 [25] Schilling JC, Mittelstedt C. Local buckling analysis of omega-stringer-stiffened composite panels using a new closed-form analytical approximate solution.Thin-Walled Structures, 2020, 147: 106534doi:10.1016/j.tws.2019.106534 [26] Xu Y, Tong Y, Liu M, et al. A new effective smeared stiffener method for global buckling analysis of grid stiffened composite panels.Composite Structures, 2016, 158: 83-91doi:10.1016/j.compstruct.2016.09.015 [27] Lu L, Meng X, Mao Z, et al. DeepXDE: A deep learning library for solving differential equations.SIAM Review, 2021, 63(1): 208-228doi:10.1137/19M1274067 [28] Tauchert TR. Large Plate Deflections, von Kármán Theory, Statical Problems//Encyclopedia of Thermal Stresses. Hetnarski RB ed. Dordrecht: Springer Netherlands, 2014: 2697-2704 [29] 沈惠申. 板壳后屈曲行为. 上海: 上海科学技术出版社, 2014Shen Huishen. Postbuckling Behavior of Plates and Shells. Shanghai: Shanghai Scientific & Technical Publishers, 2014 (in Chinese) [30] Jagtap AD, Kawaguchi K, Em Karniadakis G. Locally adaptive activation functions with slope recovery for deep and physics-informed neural networks.Proceedings of the Royal Society A:Mathematical,Physical and Engineering Sciences, 2020, 476(2239): 20200334doi:10.1098/rspa.2020.0334 [31] Chu J, Liu X, Zhang Z, et al. A novel method overcoming overfitting of artificial neural network for accurate prediction: Application on thermophysical property of natural gas.Case Studies in Thermal Engineering, 2021, 28: 101406doi:10.1016/j.csite.2021.101406 [32] 唐明健, 唐和生 基于物理信息的深度学习求解矩形薄板力学正反问题. 计算力学学报, 2022, 39(1): 120-128Tang Mingjian, Tang Hesheng. A physics-informed deep learning method for solving forward and inverse mechanics problems of thin rectangular plates.Chinese Journal of Computational Mechanics, 2022, 39(1): 120-128 (in Chinese) [33] Jiao Y, Lai Y, Li D, et al. A rate of convergence of Physics Informed Neural Networks for the linear second order elliptic PDEs.Communications in Computational Physics, 2022, 31(4): 1272-1295doi:10.4208/cicp.OA-2021-0186 [34] Krishnapriyan A, Gholami A, Zhe S, et al. Characterizing possible failure modes in physics-informed neural networks.Advances in Neural Information Processing Systems, 2021, 34: 26548-26560 [35] De Ryck T, Mishra S. Error analysis for physics-informed neural networks (PINNs) approximating Kolmogorov PDEs.Advances in Computational Mathematics, 2022, 48(6): 1-40 [36] Psaros AF, Kawaguchi K, Karniadakis GE. Meta-learning PINN loss functions.Journal of Computational Physics, 2022, 458: 111121doi:10.1016/j.jcp.2022.111121