PINN Structural Analysis

Theoretical Foundations of PINNs for Structural Analysis

Representing the Displacement Field with a Network

The structural flavour of the PINN represents the displacement field \( \mathbf{u}(\mathbf{x}) \) with a neural network and trains it to satisfy the governing equations of elasticity. Formulations fall into two broad families.

The energy form is the idea of "using the variational principle directly as the network's loss". In structural problems it often trains more stably than the strong form, which makes it worth treating as the first candidate.

How Structures Differ from Fluids

Three circumstances set structural PINNs apart from the fluid case (PINNs for fluids). (1) In most problems the solution is smooth and mild, so even a low-capacity network can represent it — a property that suits PINNs. (2) On the other hand the main battleground tends to be stress concentrations and singularities (corners, cracks), exactly where networks are weak. (3) Path-dependent nonlinearity such as plasticity and contact is fundamentally at odds with the PINN premise of representing a field by a single function, and remains at the research stage. In short, inverse problems and data fusion in linear elasticity are the practical range, while strongly nonlinear forward problems remain FEM's exclusive territory.

Where PINNs Actually Win

Numerical Methods That Make Training Work

Hard Boundary Conditions and Nondimensionalization

Here too, making the displacement boundary conditions hard (using a distance function \( \phi(\mathbf{x}) \) to construct \( \mathbf{u} = \mathbf{u}_D + \phi\,\mathcal{N} \)) is the key to training stability. And because stress (order GPa) and displacement (order mm) sit orders of magnitude apart, nondimensionalize thoroughly with a characteristic length and characteristic stress so that every output and every loss term is O(1). Skip this and only the terms with large gradients get learned, producing failures of the form "displacements match but the stresses are nonsense".

Lowering the Derivative Order with a Mixed Formulation

When differentiation noise in the stress is a problem in the strong form, a mixed formulation that carries displacement and stress as separate outputs is effective. Impose the constitutive law \( \boldsymbol{\sigma} = \mathbb{C}:\boldsymbol{\varepsilon}(\mathbf{u}) \) as an extra loss term and write equilibrium using only first derivatives of the stress output — removing the high-order derivatives cuts both automatic-differentiation cost and numerical noise, and improves the quality of the stress field. It is especially effective for plates and shells, where the strong form needs fourth derivatives.

In the Energy Form, Quadrature Decides Accuracy

The DEM loss is a domain integral of energy, and evaluating it by Monte Carlo integration (averaging over random points) passes the integration error straight through to the solution error. In practice the discipline is to (1) design the integration points on a fixed grid with Gauss-quadrature-like weights, (2) cluster integration points at stress concentrations, and (3) double the number of integration points and confirm the solution does not move — that is, run a "quadrature convergence check" in place of a mesh convergence check.

Applying It in Practice

The Winning Pattern — Material Identification from Full-Field DIC Data

1. Measure — acquire full-field displacement and strain on the specimen by DIC
2. Formulate — make the displacement network and the unknown material parameters (modulus distribution, anisotropic constants) trainable variables together. Loss = equilibrium residual + DIC agreement term + boundary conditions
3. Train — nondimensionalize, then Adam, then L-BFGS. Record the convergence history of the estimated parameters
4. Validate — agreement with DIC data on a held-out region, and a cross-check of an FEM forward analysis using the identified values against the measurement

Compared with conventional FEMU (FEM updating), the advantage is that no correspondence between the FEM mesh and the measurement points is needed, and material constants as distributions (a spatially varying Young's modulus) are handled naturally. Defect detection uses the same framework, estimating a "stiffness-reduction field" as the unknown function.

Verification Discipline — Never Skip the FEM Reference Comparison

A report on a PINN structural analysis should include (1) norm-wise error against an FEM reference solution of the same problem (separately for displacement and stress), (2) the energy balance (external work equal to internal energy), (3) error at held-out observations, and (4) the spread of results over several training runs with different random seeds. Stress in particular carries one extra order of error amplification relative to displacement, so set the tolerances on the premise that "displacement off by 1%, stress off by 10%" is normal.

A Quick Applicability Table

Frameworks and Implementation

Comparing the Implementation Routes

Implementation Skeleton (Energy Form, 2-D Plane Stress)

import torch

net = MLP(in_dim=2, out_dim=2, width=64, depth=5, act=torch.tanh)  # u(x,y)

def energy_loss(x_int, w_int, x_trac, t_bar):
    u = hard_bc(net, x_int)                  # enforce u=0 exactly via distance function
    eps = strain(u, x_int)                   # strain by automatic differentiation
    W = 0.5 * torch.sum(stress(eps) * eps, dim=1)   # strain energy density
    Pi_int = torch.sum(w_int * W)            # domain integral (designed points + weights)
    u_t = hard_bc(net, x_trac)
    Pi_ext = torch.sum(t_bar * u_t) * ds     # work of the surface traction
    return Pi_int - Pi_ext                   # quantity to minimize

The practical configuration adds nondimensionalization, the quadrature convergence check, and comparison against an FEM reference on top of this skeleton. For material identification, add a second network such as E_field = param_net(x) and train it jointly under the same loss.

Research Frontiers

Extending to Fracture and Damage — Phase-Field PINNs

Representing crack propagation with a phase field (a field of damage variable) is formally a good match for PINNs, and research on predicting crack paths with networks is active. Training is nevertheless hard because of the non-convex energy landscape, so the standard recipe is continuation learning: raise the load in stages and initialize each stage from the previous solution. Demonstrations so far are mainly 2-D on simple geometries, and practical deployment is one step further out.

Taking On Path Dependence (Plasticity, Viscoelasticity)

For history-dependent problems such as plasticity, the naive PINN of representing a field by a single function does not work, so three lines are being pursued in parallel: (1) incremental schemes that update the network at each time or load increment, (2) formulations that add internal variable fields as extra outputs, and (3) hybrids that replace only the constitutive law with a network embedded inside FEM. The closest to practice is (3); learning a constitutive law from experimental data and using it in FEM is beginning to be integrated into commercial workflows.

Placing Operator Learning and Surrogates

Operator learning (DeepONet, FNO, GNN families), which learns the map "geometry and load to stress field" from large sets of FEM results, is powerful for repeated evaluation (optimization, real-time digital twins). The division of labour is the same as for fluids: think in terms of PINNs for inverse problems with only one case of data, operator learning for fast evaluation over many cases, and Kriging for UQ with a small number of cases — a three-way split that makes it hard to choose wrongly between structures and ML.

Troubleshooting

Symptoms, Causes, and Fixes

Coexistence with FEM Is the Premise

Related: PINNs for fluid analysis, all PINN articles, Bayesian calibration (for assessing the uncertainty of an identification).