When engineers compare a finite element model of a bridge with readings from strain sensors on the real structure, the two never agree exactly. The mismatch has two different physical origins. Some of it is compatible: it could be removed by a smooth correction to how the structure deforms, as when a stiffness value or a support condition is slightly wrong. The rest is incompatible: no continuous deformation of the modeled body can produce it. Locked-in thermal strain, fabrication misfit, residual stress and slipping joints are examples. Today's Bayesian digital twins use a single smooth error term that cannot tell these two kinds apart. We propose to split the error along exactly this line, using a mesh representation that knows both where the nodes sit and which edges and faces connect them.
The construction is concrete. Fibre-optic strain gauges measure how much a short segment of material stretches. On a simplicial complex (a mesh built from vertices, edges, triangles and tetrahedra), that reading is a quantity attached to edges, not to nodes. The rigidity matrix is the linear map from node displacements to edge stretches, and it depends on both node coordinates and connectivity. It divides every possible pattern of edge stretches into two parts that do not overlap. One part comes from some displacement field. The other is the space of self-stresses, stretch patterns no displacement can produce. We put a separate Gaussian-process prior on each part. Fitted to real sensor data, the model returns the size and location of each error type, and each type points to a different engineering action.
Core Claim
A model-error prior written on the edges of the mesh, and split by the rigidity matrix into compatible and incompatible parts, should do three things that a node-level smooth error prior cannot. It should predict strain at unmonitored locations better when the true error includes locked-in strain. It should localize joint slip and residual strain. It should also say which sensor layouts can detect incompatible error at all. The last point is topological: incompatibility can only be seen when gauge paths close loops in the mesh, because readings along an open path can always be integrated into some displacement. The proposal treats simplicial representation as the core mechanism and experimental validation as the regime where it must succeed, on real gauges with real noise.
Relation to Prior Work
The Bayesian side comes from the statistical finite element method. The Statistical Finite Element Method treats the solver output as a probabilistic prior and adds a Gaussian-process discrepancy term for missing physics. Digital twinning of self-sensing structures using the statistical finite element method applies it to 108 fibre Bragg grating sensors on Intersection Bridge 20A. That study reconstructs strain on an uninstrumented beam and keeps accuracy high when the sensors are reduced from 40 to 10. We keep this inference machinery. We change what the discrepancy is defined on, from smooth fields over space to structured patterns on mesh edges, so it is no longer one smooth field.
The simplicial side supplies the operators. Continuous Simplicial Neural Networks (COSIMO) runs heat diffusion over Hodge Laplacians, the edge-level analogues of the graph Laplacian. It learns separate diffusion times for the "down" direction through faces and the "up" direction through cofaces, and it uses truncated eigendecompositions to control the cubic cost. We reuse the edge heat kernel as a smoothness prior inside the incompatible subspace. We also reuse the truncation strategy, and we treat the diffusion times as hyperparameters fitted by marginal likelihood.
Clifford Group Equivariant Simplicial Message Passing Networks motivates the claim that keeping oriented geometric information on higher simplices beats collapsing it to scalars. For example, it reports a mean squared error of 0.002 against 0.011 for an equivariant graph network on convex-hull volume prediction. Neither simplicial paper touches physical sensors or constitutive mechanics. We use no deep network in the core method, because the question here is the structure of the error, not speed.
Neural mesh solvers, Learning Mesh-Based Simulation with Graph Networks and Physics-informed MeshGraphNets, pass messages only along edges and are evaluated only against simulated data. They constrain our scope: they could replace the finite element prior later, but they do not address error structure.
Effect of Reinforcing Bar on Rayleigh Wave Propagation on Concrete Structures and Experimental and Numerical Study of Static Behavior of Precast Segmental Hollow Bridge Piers motivate the incompatible term. In both, unmodelled internal discontinuities dominate the mismatch between measurement and simulation: rebar scattering in one, joint slip at the pier base in the other.
Research Program
Let $R$ be the rigidity matrix of the mesh, of size edges by three times vertices. Let $\Pi$ be the orthogonal projector onto the null space of $R^\top$, the self-stress space. Let $L_1$ be the edge Hodge Laplacian. The edge-level error $\delta$ in strain gets the prior
$$ \delta \sim \mathcal{N}\!\left(0,\; \sigma_c^2\, R K_u R^\top \;+\; \sigma_i^2\, \Pi\, e^{-t L_1}\, \Pi\right), $$
where $K_u$ is a smooth covariance over node displacements, $\sigma_c$ and $\sigma_i$ scale the compatible and incompatible parts, and $t$ is a diffusion time. A sensor reading is a weighted sum of edge stretches along its gauge path plus noise, $y = H(\varepsilon_{\text{FEM}} + \delta) + \eta$. Here $H$ maps edge stretches to gauge readings, $\varepsilon_{\text{FEM}}$ is the finite element strain, and $\eta$ is sensor noise. Because the model is linear and Gaussian, the posteriors over both error parts and over unmonitored strains come out in closed form.
The first term alone, with $K_u$ smooth in space, is essentially what statFEM already does. The second term is the new ingredient. Its effect on the data is bounded by how much the gauge paths, through $H$, overlap with the self-stress space. That overlap gives a computable identifiability score for any sensor layout before anyone installs a gauge.
The work has four parts.
- Implementation. We build $R$, $\Pi$ and a truncated $e^{-tL_1}$ on tetrahedral and shell meshes. Projections use sparse QR factorization and the heat kernel uses Lanczos iteration, so we avoid full eigendecompositions.
- Identifiability theory. We prove that gauge layouts with no mesh cycles have zero information about $\sigma_i$.
- Identifiability as a function of layout. We characterize identifiability as a function of rosette density and loop closure.
- Validation. We test on progressively more realistic data, described below.
Baselines
The naive reference is the uncorrected finite element model, with gauge noise estimated from repeated static readings. Any method must beat its held-out strain error to be worth deploying.
The strongest existing competitor is statFEM as in Digital twinning of self-sensing structures using the statistical finite element method. Its discrepancy is a smooth Gaussian process that is then mapped to strain. It was built for and validated on precisely this sensor type and structure. We found no reported numerical posterior error or interval coverage on Lacuna for statFEM on Bridge 20A; the paper is summarized only as accurate, with calibrated 95% intervals, when sensors are reduced from 40 to 10. We will therefore re-implement it on the same meshes, finite element priors and sensor splits as our method. Its kernel hyperparameters get the same marginal-likelihood optimization budget, with an equal number of restarts. Beating it means two things on held-out sensors: a lower negative log predictive density, and 95% coverage within two points of nominal. On tests where the true error is purely compatible, we must at least tie it, otherwise the extra term is costing us.
A purely data-driven reference is Gaussian-process kriging of the sensor strains with no finite element prior. It tests whether physics helps at all.
Two ablations isolate the key ingredient. The first replaces the rigidity-based split with the plain Hodge decomposition of $L_1$. That decomposition uses connectivity only and ignores coordinates, so this ablation shows whether geometry matters or only topology. The second keeps both terms but uses an isotropic edge kernel in place of $\Pi$, which shows whether the separation itself matters or only the extra capacity.
Why not just…?
Why not just use statFEM with a richer node-level kernel? This is the real threat. A flexible enough smooth error field mapped to strain can mimic sharp local gradients near a slipping joint. In fact, on a continuous mesh, slip appears partly as a large local but still compatible gradient. What it cannot represent is genuine self-stress, such as thermal or misfit eigenstrain closed around loops. The injected-error study below separates these cases. If statFEM with a nonstationary kernel matches us on the eigenstrain cases as well, the incompatible term adds nothing and we will report that.
Why not just interpolate the sensors directly? Kriging along a single instrumented flange will probably interpolate well. It has no route to transfer across beams. The East-to-West Bridge 20A protocol tests exactly that transfer, and so do the plate experiments with held-out regions.
Why not just add the missing physics to the finite element model, for example cohesive joint elements as in the pier study? When the mechanism and its location are known, that is the better answer. The incompatible posterior is meant for the case where they are not. Its localization tells the engineer where to add such elements. We test this directly by checking whether the localization finds the joint we deliberately loosen.
Why not a neural surrogate such as MeshGraphNets? That would speed up the prior, but it leaves the error structure untouched, and those models have been validated only against simulations. It is orthogonal to this proposal and can be swapped in later.
Proposed Experiments
Stage one: synthetic data with known truth. Using a finite element "truth" model, we inject three kinds of error into otherwise matching models. The first is a stiffness error, which is compatible. The second is a thermal eigenstrain gradient, which is incompatible. The third is joint slip modelled with cohesive elements, following the pier study. We sample gauge layouts ranging from open single lines to closed rosette loops. The predicted outcome is that our method ties statFEM on the stiffness error, beats it on eigenstrain, and fails to identify the incompatible part on open layouts, exactly as the identifiability score predicts. That last prediction is itself a check of the theory.
Stage two: a new laboratory specimen. We build a bolted aluminium plate assembly instrumented with strain-gauge rosettes and run controlled heating and staged loosening of the bolts. Full-field digital image correlation, a camera-based measurement of surface displacement, is held out as ground truth. The decisive check is whether the posterior incompatible magnitude rises with bolt loosening and concentrates at the loosened joint, while the compatible part stays flat.
Stage three: the operational bridge. If the Intersection Bridge 20A data can be obtained, we repeat the published East-to-West and 40-to-10 sensor protocols. We report held-out strain error, negative log predictive density and interval coverage against our re-run of statFEM.
The proposal fails if the method shows no gain over statFEM on the incompatible-error cases in stage one, or if it cannot localize the loosened joint in stage two.
References
- The Statistical Finite Element Method
- Digital twinning of self-sensing structures using the statistical finite element method
- Continuous Simplicial Neural Networks
- Clifford Group Equivariant Simplicial Message Passing Networks
- Learning Mesh-Based Simulation with Graph Networks
- Physics-informed MeshGraphNets (PI-MGNs): Neural finite element solvers for non-stationary and nonlinear simulations on arbitrary meshes
- Effect of Reinforcing Bar on Rayleigh Wave Propagation on Concrete Structures
- Experimental and Numerical Study of Static Behavior of Precast Segmental Hollow Bridge Piers