Physics-informed deep learning for incompressible laminar flows

Physics-informed deep learning has drawn tremendous interest in recent years to solve computational physics problems, whose basic concept is to embed physical laws to constrain/inform neural networks, with the need of less data for training a reliable model. This can be achieved by incorporating the residual of physics equations into the loss function. Through minimizing the loss function, the network could approximate the solution. In this paper, we propose a mixed-variable scheme of physics-informed neural network(PINN) for fluid dynamics and apply it to simulate steady and transient laminar flows at low Reynolds numbers. A parametric study indicates that the mixed-variable scheme can improve the PINN trainability and the solution accuracy. The predicted velocity and pressure fields by the proposed PINN approach are also compared with the reference numerical solutions. Simulation results demonstrate great potential of the proposed PINN for fluid flow simulation with a high accuracy.     

基于物理信息神经网络的功能梯度材料稳态/瞬态热传导分析

采用物理信息神经网络PINN(Physics-informed Neural Networks)求解稳态和瞬态功能梯度材料(FGMs)热传导问题。该方法利用控制方程、边界及初始条件的残差构造损失函数,在无任何响应数据的情况下得到了更具泛化能力的神经网络模型,同时避免了传统数值方法在求解计算力学问题时所需的微分、积分公式推导以及繁重的建模和划分网格等前处理工作。本文探究了PINN及其域分解的扩展物理信息神经网络XPINN(eXtended Physics-informed Neural Networks)在求解稳态和瞬态FGMs热传导问题时的适用性,讨论了网络结构对预测结果的影响。研究结果表明,PINN/XPINN在解决几何复杂的稳态和瞬态FGMs热传导问题时仍具有较高的可靠性和简洁的求解流程,同时,为极端环境下求解复杂多场耦合和夹杂等问题提供了新思路。     

多晶金属材料的三维组集式弹塑性本构模型

本文从分析多晶金属材料的细观组织在弹塑性变形中的贮能和耗能机制入手,提出一个三维组集式弹塑性本构模型。该模型将材料单元抽象成沿三维空间各方向均匀分布组件的集合体,方向组件反映材料的细观性态并在宏观上协调变形,所有方向组件的内力总效应构成宏观应力。文中导出了显式的弹塑性本构关系,并与Budiansky的复杂加载试验结果及其它塑性模型进行了对比。     

FD-PINN:频域物理信息神经网络

物理信息神经网络(physics-informed neural network, PINN)是将模型方程编码到神经网络中,使网络在逼近定解条件或观测数据的同时最小化方程残差,实现偏微分方程求解.该方法虽然具有无需网格划分、易于融合观测数据等优势,但目前仍存在训练成本高、求解精度低等局限性.文章提出频域物理信息神经网络(frequency domain physics-informed neural network, FD-PINN),通过从周期性空间维度对偏微分方程进行离散傅里叶变换,偏微分方程被退化为用于约束FD-PINN的频域中维度更低的微分方程组,该方程组内各方程不仅具有更少的自变量,并且求解难度更低.因此,与使用原始偏微分方程作为约束的经典PINN相比, FD-PINN实现了输入样本数目和优化难度的降低,能够在降低训练成本的同时提升求解精度.热传导方程、速度势方程和Burgers方程的求解结果表明, FD-PINN普遍将求解误差降低1～2个数量级,同时也将训练效率提升6～20倍.     

基于深度学习和细观力学的颗粒材料本构关系研究

颗粒材料的本构关系对岩土工程等众多领域至关重要. 不同于传统的唯象本构理论, 本文基于机器学习模型探索了一种细观力学理论指导下的数据驱动型颗粒材料本构关系预测方法. 根据Vogit均质化假设, 建立了小应变条件下颗粒材料应力?应变解析关系, 此关系唯一地确定了一组与颗粒材料本构行为相关的细观组构变量. 这些变量与反应颗粒材料宏观性质的主应力和主应变信息通过一系列离散元三轴压缩数值试验获得. 考虑到细观组构变量为内变量, 不能直接作为本构模型的输入. 本文基于有向图方法将颗粒材料微观结构信息隐式地包含在应力?应变的预测当中, 并采用门控循环单元(GRU)循环神经网络作为基础深度学习模型描述有向图中结点之间的映射关系. 通过将有向图从目标节点沿源节点展开, 整个应力?应变预测模型可由两个神经网络分别训练并组装而成. 将训练后的深度学习模型在全新的数据集上进行测试, 结果表明该训练策略能有效捕捉到颗粒材料在常规三轴任意加卸载, 等中主应力系数b的真三轴加载, 和等平均有效应力p的真三轴加卸载等复杂多轴加载工况下的应力?应变响应关系, 模型具有良好的内插和外推预测能力. 考虑到深度学习模型捕捉颗粒材料力学响应的能力及其开放式学习的特点, 充分结合数据驱动方法和理论本构模型可能是颗粒材料本构研究的一个重要方向.      

Heuristic search for a predictive strain-energy function in nonlinear elasticity

In this work, a new, quasi-structural model – bootstrapped eight-chain model – is proposed as a modification to the strain energy of eight-chain model [Arruda, E.M., Boyce, M.C., 1993. A three-dimensional constitutive model for the large stretch behaviour of rubber elastic materials. J. Mech. Phys. Solids 41, 389—412] that invokes the Langevin chain statistics. This development has been led to by our heuristic search into how the strain energy of eight-chain model may be adapted in order to account better for the mechanical behaviour of elastomeric materials in both linear and nonlinear elastic regimes [Treloar, L.R.G., 1944. Stress–strain data for vulcanised rubber under various types of deformation. Trans. Faraday Soc. 40, 59–70]. The eight-chain model appears to produce very similar results in predicting biaxial stress to those of a first stretch-invariant model that gives a good fit in uniaxial extension and, thus, it is shown that the former can not be significantly enhanced within the limitation of the latter. Evaluation of predictive capability for an additive invariant-separated form of strain energy shows that an explicit inclusion of a second stretch-invariant function would not work and that any thus added term ought to be dependent on both the first and second stretch-invariants of deformation tensor, and hints that an improvement is possibly needed at low strain. The composite and filament models [Miroshnychenko, D., Green, W.A., Turner, D.M., 2005. Composite and filament models for the mechanical behaviour of elastomeric materials. J. Mech. Phys. Solids 53 (4), 748–770] have their strain-energy functions in that suggested form and cope very well with predicting the experimental data of Treloar (1944). We use the form of strain energy for the filament model, that proved to be successful, to bootstrap the strain energy of eight-chain model in order to improve the performance of the latter at low strain. Thus, we derive a new model – bootstrapped eight-chain model – that requires only two material parameters – a rubber modulus and a limiting chain extensibility. The proposed model is quasi-structural due to bootstrapping and it retains the best traits and corrects the faults of the eight-chain model, conforming more closely to the classical experimental data of Treloar (1944).     

Reversed planar elongation of soft polymeric networks

The newly developed planar elongation fixture, designed as an add-on to the filament stretch rheometer, is used to measure reversible large amplitude planar elongation on soft elastomers. The concept of this new fixture is to elongate an annulus, by keeping the perimeter constant. The deformation on the cylindrical probe is measured using digital imaging, and it is found that the diameter only changes a few percent compared with the initial diameter. Additionally, it is found that a time-strain separable K-BKZ model based on a Doi?CEdwards (DE) network strain (without independent alignment) captures the experimental data well. In particular, it is observed that it reproduces the deformation on the cylindrical probe, and this observation confirms our previous statement that the deformation on the cylindrical probe is highly sensitive towards the choice of strain tensor. The stress itself is well described by both the modified Lodge and the DE network strain. The amount of work needed during the deformation is calculated, and it is illustrated how the sample itself contributes with work upon flow reversal. As the stress is time-strain separable the energy loss originates from linear viscoelastic relaxation even in the nonlinear regime, and the loss can be determined solely from linear viscoelastic measurements.     

考虑应变率效应的混凝土随机损伤本构模型研究

混凝土材料组分复杂且具有随机分布的特点,其受力力学行为不可避免地存在非线性和随机性.同时,在动力荷载作用下,混凝土材料具有不可忽视的率敏感性.为了综合反映混凝土受力力学行为中的非线性、随机性与率敏感性,本文从对材料的纳-微观裂纹扩展分析入手,引入速率过程理论描述纳观裂纹的扩展速率,并研究了对应的能量耗散过程.在此基础上通过裂纹层级模型将纳观分析推演到微观尺度,建立了微观能量耗散的基本表达式.进而与微-细观随机断裂模型相结合,形成了混凝土纳-微-细观随机损伤本构模型.同时,基于速率相关势垒的分析,揭示了动力强度的提高源自加载速率和原子键断裂速率的竞争机制.据此,假定微裂纹间相互作用与应变率比值的相关关系以建立微弹簧能量耗散速率与应变率的联系,实现了从静力本构模型向动力本构模型的扩展.数值算例表明,建议模型能够同时反映混凝土材料力学行为中的非线性、随机性和率敏感性.最后通过与相关试验结果的对比,验证了建议模型的正确性.     

数据驱动梯度结构材料弹塑性本构

梯度结构材料因其优异的力学性能被广泛应用于工程结构中.论文整合塑性理论和人工神经网络技术,发展了一种构建梯度结构材料弹塑性本构模型的新方法.该方法基于梯度结构材料不同位置的微结构,构建不同代表性体积单元,进而生成应力应变数据,应用生成的数据训练人工神经网络,建立基于神经网络的材料本构模型.应用该方法,论文开展了针对实际...     

基于广义应变梯度理论的纳米梁挠曲电效应研究

挠曲电效应是应变梯度与电极化的耦合,它存在于所有的电介质材料中。在纳米电介质结构的挠曲电效应研究中,应变梯度弹性对挠曲电响应的影响一直以来被低估甚至被忽略了。根据广义应变梯度理论,应变梯度弹性中独立的尺度参数只有三个,而文献中所采用的一个或两个尺度参数的应变梯度理论只是它的简化形式。基于该理论,论文建立了考虑广义应变梯度弹性的三维电介质结构的理论模型,并以一维纳米梁为例研究了其弯曲问题的挠曲电响应及其能量俘获特性。结果表明,纳米梁的挠曲电响应存在尺寸效应,并且弹性应变梯度会影响结构挠曲电的尺寸效应,特别是当结构的特征尺寸低于尺度参数时。论文的工作为更进一步理解纳米尺度下的挠曲电机理和能量俘获特性提供理论基础和设计依据。     

金属材料的率-温耦合响应与动态本构关系综述

金属材料在冲击、爆炸等高应变率加载下的塑性流动行为具有不同于静载下的率-温耦合性和微观机制。航空航天、航海、能源开采、核工业、公共安全、灾害防治等方面的金属结构设计与性能评估需要进行大量的动载实验和数值模拟,建立准确的材料动态本构模型是结构数值模拟可靠性的基础和关键。本文中,总结了金属材料的率-温耦合变形行为及内在机理,回顾了金属动态本构关系研究的起源与发展脉络,分别针对唯象模型、具有物理基础的模型和人工神经网络模型进行了详细介绍和横向比较。唯象模型和人工神经网络模型分别因易应用和高预测精度而受到青睐,基于物理概念的宏观连续介质模型可以描述体现内部演化的真实物理量,从而涵盖更大的应变范围,更好地反映应变率、温度和应变的影响机制。     

基于数字散斑相关实验测量的材料构型力的计算方法

材料构型力学主要研究材料中的缺陷(夹杂、空穴、位错、裂纹、塑性区等)的构型(形状、尺寸和位置)改变时,所引起的系统自由能的变化。本研究将基于数字散斑相关技术,实验测量材料试件的位移场分布,随后通过材料构型力的定义式,计算求得弹塑性材料中缺陷构型力的分布。其方法概括如下:位移场通过数字图像相关技术测得;应变及位移梯度场利用三次样条拟合获得;线弹性材料应力通过简单线弹性本构方程获取,而塑性材料的表面应力场通过Ramberg-Osgood本构方程计算求得;弹塑性应变能密度分布则由应力-应变曲线数值积分获得。该方法对普通弹性材料或者弹塑性材料均适用,可以用于各种不同的缺陷及缺陷群的材料构型力测量。     

Solving Huxley equation using an improved PINN method

Although many effective methods for solving partial differential equations (PDEs) have been proposed, there is no universal method that can solve all PDEs. Therefore, solving partial differential equations has always been a difficult problem in mathematics, such as deep neural network (DNN). In recent years, a method of embedding some basic physical laws into traditional neural networks has been proposed to reveal the dynamic behavior of equations directly from space-time data [i.e., physics-informed neural network (PINN)]. Based on the above, an improved deep learning method to recover the new soliton solution of Huxley equation has been proposed in this paper. As far as we know, this is the first time that we have used an improved method to study the numerical solution of the Huxley equation. In order to illustrate the advantages of the improved method, we use the same network depth, the same hidden layer and neurons contained in the hidden layer, and the same training sample points. We analyze the dynamic behavior and error of Huxley’s exact solution and the new soliton solution and give vivid graphs and detailed analysis. Numerical results show that the improved algorithm can use fewer sample points to reconstruct the exact solution of the Huxley equation with faster convergence speed and better simulation effect.

     

Learning material law from displacement fields by artificial neural network

The recently developed data-driven approach can establish the material law for nonlinear elastic composite materials(especially newly developed materials) by the generated stress-strain data under different loading paths(Computational Mechanics, 2019). Generally, the displacement(or strain) fields can be obtained relatively easier using digital image correlation(DIC) technique experimentally, but the stress field is hard to be measured. This situation limits the applicability of the proposed data-driven approach. In this paper, a method based on artificial neural network(ANN) to identify stress fields and further obtain the material law of nonlinear elastic materials is presented, which can make the proposed data-driven approach more practical. A numerical example is given to prove the validity of the method. The limitations of the proposed approach are also discussed.     

The evolution of laminates in finite crystal plasticity: a variational approach

The analysis and simulation of microstructures in solids has gained crucial importance, virtue of the influence of all microstructural characteristics on a material’s macroscopic, mechanical behavior. In particular, the arrangement of dislocations and other lattice defects to particular structures and patterns on the microscale as well as the resultant inhomogeneous distribution of localized strain results in a highly altered stress–strain response. Energetic models predicting the mechanical properties are commonly based on thermodynamic variational principles. Modeling the material response in finite strain crystal plasticity very often results in a non-convex variational problem so that the minimizing deformation fields are no longer continuous but exhibit small-scale fluctuations related to probability distributions of deformation gradients to be calculated via energy relaxation. This results in fine structures that can be interpreted as the observed microstructures. In this paper, we first review the underlying variational principles for inelastic materials. We then propose an analytical partial relaxation of a Neo-Hookean energy formulation, based on the assumption of a first-order laminate microstructure, thus approximating the relaxed energy by an upper bound of the rank-one-convex hull. The semi-relaxed energy can be employed to investigate elasto-plastic models with a single as well as multiple active slip systems. Based on the minimization of a Lagrange functional (consisting of the sum of energy rate and dissipation potential), we outline an incremental strategy to model the time-continuous evolution of the laminate microstructure, then present a numerical scheme by means of which the microstructure development can be computed, and show numerical results for particular examples in single- and double-slip plasticity. We discuss the influence of hardening and of slip system orientations in the present model. In contrast to many approaches before, we do not minimize a condensed energy functional. Instead, we incrementally solve the evolution equations at each time step and account for the actual microstructural changes during each time step. Results indicate a reduction in energy when compared to those theories based on a condensed energy functional.     

The complete synchronization of Morris–Lecar neurons influenced by noise

The influence of noise on the complete synchronization in a Morris–Lecar (ML) model neuronal system is studied in this work. Two individual ML neurons with different initial conditions can discharge completely synchronously when the noise intensity is large enough, and for a smaller reversal potential (V _Ca), the uncoupled neuronal system could be induced to a complete synchronization state under smaller noise intensity. Two coupled ML neurons could be synchronized under very small noise intensity even in the case of weak coupling, the synchronization characteristics of the two coupled neurons are discussed by analyzing the Similarity Function (S(0)) of their membrane potentials, which proves that noise can promote the complete synchronization. The critical noise intensity (D _j ) to induce complete synchronization in coupled ML neurons will decrease with the increase of V _Ca. This result is helpful to study the synchronization and the code of a neural system.     

Limit point instability in pressurization of anisotropic finitely extensible hyperelastic thin-walled tube

Mechanical responses of materials undergoing large elastic deformations can exhibit a loss of stability in several ways. Such a situation can occur when a thin-walled cylinder is inflated by an internal pressure. The loss of stability is manifested by a non-monotonic relationship between the inflating pressure and internal volume of the tube. This is often called limit point instability. The results, known from the literature, show that isotropic hyperelastic materials with limiting chain extensibility property always exhibit a stable response if the extensibility parameter of the Gent model satisfies J_m<18.2. Our study investigates the same phenomenon but for tubes with anisotropic form of the Gent model (finite extensibility of fibers). Anisotropy, used in our study, increases the number of material parameters the consequence of which is to increase degree of freedom of the problem. It will be shown that, in stark contrast to isotropic material, the unstable response is predicted not only for large values of J_m but also for J_m≈1 and smaller, and that the existence of limit point instability significantly depends on the orientation of preferred directions and on the ratio of linear parameters in the strain energy density function (this ratio can be interpreted as the ratio of weights by which fibers and matrix contribute to the strain energy density). Especially tubes reinforced with fibers oriented closely to the longitudinal direction are susceptible to a loss of monotony during pressurization.     

红板岩材料隐式本构模型的建立及其应用研究

针对红板岩材料在岩土工程中所表现的大量模糊的和不确定的因素等特点,基于人工神经网络的学习能力,借助于室内岩石力学试验,进行了对该材料的力学本构特性进行了神经网络模拟研究,提出了隐式本构模型的思想和方法,并通过该方法对该岩石的流变试验结果进行学习,获得了以网络权值结构保存的力学特性知识,由此得到了表征红板岩应力应变本构关系的隐式本构模型。应用结果表明,该方法对岩土类材料本构关系的模拟研究具有很好的应用前景。     

Finite element analysis of V-ribbed belts using neural network based hyperelastic material model

A three-dimensional finite element model was built to study V-ribbed belt pulley contact mechanics. The model consists of a pulley and a segment of V-ribbed belt in contact with the pulley. A material model for the belt, including the rubber compound and the reinforcing cord is developed. Rubber is modeled as hyperelastic material. The hyperelastic strain energy function is approximated by neural network trained by rubber test data. Reinforcing cord is modeled as elastic rebar. The material model developed is implemented in the commercial finite element code ABAQUS to simulate the V-ribbed belt-pulley system. A study is then conducted to investigate the effect of belt pulley system parameters on the contact mechanics. The effects of temperature and aging on belt materials are also investigated. The information gained from the analysis can be applied to optimize V-ribbed belt and pulley design.