Stress Desaturation: Why Mean-Seeking Stress Reconstructors Underestimate Failure, and a Quantized Fix

When a neural network must guess a quantity that has several plausible values, and it is trained with squared error, it learns to output the average of those values. In image colorization this is a familiar failure. A gray shirt could be red or blue, so the network paints it a muddy brown. Colorization researchers call this desaturation. They fixed it by turning color prediction into classification over a grid of quantized color bins. This note argues that the same failure affects neural reconstruction of mechanical stress from sparse or ambiguous measurements, and that there it has a sharper consequence than dull pictures. Engineers judge failure with criteria such as von Mises stress, which, like color saturation, is the length of a vector. Length is a convex function, and the length of an average is never larger than the average of lengths. A network that outputs the conditional mean of the stress tensor will therefore predict lower failure indicators than the truth whenever the measurements are ambiguous. The error always points in the unsafe direction.

The proposal makes three claims and tests each one. First, peak-stress underestimation in neural stress models has two different causes that the literature runs together. Forward surrogates, which map a known geometry and load to a unique stress field, have no ambiguity, so whatever peak error they show comes from class imbalance and smoothing, not from mean-seeking. Inverse reconstruction from sparse sensors or photoelastic images is genuinely ambiguous, and there mean-seeking produces a systematic bias whose size we can predict. Second, physics losses cannot remove this bias in linear elasticity. Third, a per-point categorical distribution over quantized principal stresses, trained with class rebalancing, removes the bias, and it stays compatible with physics losses.

Core Claim

Let $\sigma$ be the stress tensor at a point, $x$ the available measurements, and $f$ a convex failure measure, for example plane-stress von Mises $f(\sigma)=\sqrt{\sigma_1^2-\sigma_1\sigma_2+\sigma_2^2}$ with principal stresses $\sigma_1,\sigma_2$. A model trained with squared error converges to $\hat\sigma(x)=\mathbb{E}[\sigma\mid x]$. Jensen's inequality then gives

$$f(\mathbb{E}[\sigma\mid x]) \;\le\; \mathbb{E}[f(\sigma)\mid x],$$

and the gap between the two sides grows with how spread out $p(\sigma\mid x)$ is. Equality holds only when the conditional distribution collapses to a point, which is the forward-surrogate case. We call this gap stress desaturation. It is the same inequality that makes averaged chroma, the length of the color vector $(a,b)$, come out grayer than real chroma in colorization (Colorful Image Colorization).

The second claim follows from linearity. Static linear-elastic equilibrium, $\nabla\cdot\sigma + b = 0$ with traction boundary conditions, is a linear constraint. The average of two admissible stress fields is therefore also admissible, which means the set of physically valid fields is closed under averaging. A physics-informed residual loss gives the collapsed mean field a residual of zero, so it cannot penalize it. This is a structural reason why physics-informed training, such as Physics Informed Neural Network for Dynamic Stress Prediction, helps with smoothness but cannot be expected to fix ambiguity-driven peak loss.

The same linearity also resolves a tension the literature raises, namely that discrete outputs seem to break the derivatives that physics losses need. If the network outputs a categorical distribution per point, its expected tensor field is differentiable and should itself satisfy equilibrium. The physics residual can be applied to that expected field, while failure indicators are read from the full distribution as $\mathbb{E}[f(\sigma)\mid x]$ rather than $f(\mathbb{E}[\sigma\mid x])$.

Relation to Prior Work

The method is borrowed from colorization. Colorful Image Colorization recasts color prediction as 313-way classification over the $ab$ plane. It adds rebalancing weights that boost rare saturated bins and an annealed-mean decoder that sharpens the distribution with a temperature before averaging. CT2 adds a module that masks color tokens which are impossible given the input brightness. We reuse all three ideas, with one direct analogy: the principal-stress plane $(\sigma_1,\sigma_2)$ plays the role of the $ab$ plane, and von Mises stress plays the role of chroma. We also borrow the masking idea. Bins can be ruled out by known physical bounds, for instance bins with tensile stress on a surface known to be in compression.

The closest bridge to physical fields is ColorNeRF. It shows that fitting a continuous 3D field to monochrome views with squared error falls into a gray trap, and that injecting color through quantized classification escapes it. ColorNeRF works only on optical color, has no physics constraints, and gives no account of which downstream quantities the trap biases. Our contribution is the claim that the bias is directional for convex safety criteria, plus a test of that claim.

Two papers motivate the proposal. Neuro-DynaStress reports that its convolutional-recurrent surrogate underestimates peak stresses at 90-degree corners and when distributions are non-Gaussian, and it suggests pre-smoothing the corners. PINN-Stress reports that physics losses improved peak accuracy but that error rose to 16% on unseen boundary conditions. Both are forward problems, so our framework predicts that quantization should not be their main fix. The forward setting becomes our negative control.

Several papers constrain the design. PalSB shows that diffusion models reconstructing fields from sparse sensors produce sharp but physically invalid fields unless their generation trajectories are fine-tuned against equation residuals. ROBIN shows that diffusion in solid mechanics needs a special pipelined schedule to be fast enough to use. Together they mark the expensive generative alternative we must beat or match. NeST reconstructs 3D stress tensors from photoelastic images using a coordinate network and a differentiable optical forward model with squared-error fitting. Photoelastic data constrain only the difference of principal stresses, wrapped by fringe order, and the principal directions only up to symmetry. That makes NeST's setting a natural, strongly ambiguous testbed. On the geological side, the Precaspian low-angle fracture study shows that whether rock fails depends on local stress exceeding Coulomb-Mohr or Griffith thresholds. This is why averaging out peaks is consequential and why smoothing geometry is not a fix. We treat paleostress inversion as motivation only and do not evaluate on it.

Research Program

We will build one reconstruction network with interchangeable output heads, so that the only variable is how the network represents uncertainty about stress at each point.

The categorical head predicts, at each point, a distribution over a 2D grid of $(\sigma_1,\sigma_2)$ bins. The grid is spaced logarithmically in radius, so that the heavy high-stress tail gets resolution. A second small categorical head predicts the principal direction over 12 bins on $[0,\pi)$. Training uses rebalanced cross-entropy, with weights computed as in Zhang et al.: inverse bin frequency, smoothed and mixed with a uniform distribution. The physics residual is applied to the expected tensor field, which is assembled from the expected principal values and direction. At test time the head reports three things. It gives an annealed-mean field for visualization, $\mathbb{E}[f(\sigma)\mid x]$ for failure assessment, and a per-point exceedance probability $P(f(\sigma)>\tau\mid x)$ for an allowable stress $\tau$.

Because we design the inverse tasks ourselves, we can control how ambiguous they are. For each test case we can estimate the true conditional spread by solving for many stress fields that are consistent with the same observations. This lets us check whether the measured peak shortfall of squared-error models follows the Jensen gap predicted from that spread.

Baselines

The naive reference point is a squared-error regressor using the same backbone. For the forward control, the strongest published result on the gusset-plate data is Neuro-DynaStress at 2.3% mean relative percentage error on validation (Neuro-DynaStress). On the expanded mesh data, PINN-Stress reports 1.5% error on unseen geometries, 4.2% on unseen load sequences, and 16% on unseen boundary conditions (PINN-Stress). We will reproduce a physics-informed squared-error model in the PINN-Stress style on our regenerated data. Beating it in the forward setting is not the goal. Matching its mean-field error while not losing peak accuracy is the bar.

The strongest competitor on the inverse question is a diffusion reconstructor of the PalSB type, a model that generates full fields from sparse sensor readings and is then fine-tuned with equation residuals (PalSB). It escapes mean collapse by drawing samples. We found no reported result on Lacuna for PalSB on stress fields, and no numeric result for NeST on 3D-TSV. Both will therefore be re-run under our protocol. Each method gets the same training data, the same backbone parameter count where possible, and the same hyperparameter search budget, measured in GPU-hours. Failure statistics for the diffusion model will be estimated from its samples at matched wall-clock inference time. Beating it means matching its exceedance-probability calibration at a fraction of its inference cost, or beating its calibration at equal cost.

Two further continuous baselines test whether quantization is necessary. One is a heteroscedastic Gaussian head, which predicts a mean and a variance per point. The other is a mixture density network, which predicts a small mixture of Gaussians per point and is the strongest continuous alternative. Two ablations isolate the key ingredients. In the first, we remove rebalancing from the categorical head. In the second, we keep the categorical head but read failure from $f$ of the annealed mean instead of from the distribution, which tests whether the decoding rule, rather than the representation, carries the gain.

Why not just…?

Why not just use L1, Huber, or peak-weighted squared error? For symmetric bimodal ambiguity, such as a stress sign that cannot be determined from the data, the median sits at the same interior point as the mean. Zhang et al. report that L1-trained colorization remained desaturated. Any single-point estimator also still suffers the Jensen gap on convex criteria. Peak-weighted regression is a serious threat in the forward setting, where there is no ambiguity, and our negative control is designed to show that it suffices there. It is included as a baseline in every inverse task.

Why not just a mixture density network? This is the most credible threat. It represents multimodality continuously and is differentiable. Our argument for bins is empirical rather than principled. Bins are nonparametric, handle heavy tails through log spacing, and support rebalancing directly, whereas mixture networks are known to be awkward to train when components collapse. The head-to-head comparison across ambiguity levels decides this. If the mixture network matches the categorical head on calibration and peak ratio, we will report that the claim to emphasize is decoding failure from a distribution, not quantization itself.

Why not just diffusion? Diffusion samples do escape the mean, but PalSB shows they need physics fine-tuning to avoid invalid fields, and ROBIN shows their cost needs engineering to become practical. Our comparison is at matched inference cost. If diffusion wins at equal cost, the categorical head's case reduces to speed, and we will say so.

Why not just add a physics-informed loss? By the convexity argument above, the equilibrium residual of the mean field is zero in linear elasticity, so this loss cannot see the collapse. We test this directly. If adding the residual loss to the squared-error model fails to change the peak ratio in the inverse tasks while it does improve smoothness, the argument is confirmed.

Why not smooth the corners, as Neuro-DynaStress suggests? Smoothing changes the stress concentration factor. That concentration factor is exactly what failure criteria act on, so smoothing removes the phenomenon the model is supposed to detect.

Proposed Experiments

We will regenerate the gusset-plate data following the protocol described for Neuro-DynaStress: 1,024 pentagonal geometries, randomized sinusoidal loads at 1 to 3 Hz and 2 to 10 kN, and four boundary conditions, solved with a finite-element solver. We use static snapshots to keep the linear-equilibrium argument exact. The forward task maps known geometry and load to stress. The inverse task gives the network only $k$ strain-gauge readings, with $k$ swept from dense to very sparse, while load direction and boundary condition are hidden. This raises ambiguity in a controlled way. The third task uses 3D-TSV with the NeST optical forward model: the network reconstructs stress from simulated polariscope images, with the number of viewing axes swept.

The metrics are relative error of the mean field, the ratio of predicted to true peak von Mises stress in the top 1% of points, Brier score and reliability curves for exceedance probability, the continuous ranked probability score of the predicted per-point distribution, and the equilibrium residual of the expected field.

Three checks decide the outcome. In the forward task, the categorical head should give no peak-ratio advantage over peak-weighted regression. If it gives a large one, our account of the cause is wrong. In the inverse tasks, the squared-error model's peak-ratio shortfall should grow as $k$ decreases and should track the Jensen gap estimated from sampled consistent fields, and the categorical head should keep the peak ratio near one with calibrated exceedance probabilities. Across all inverse settings, adding the physics residual to the squared-error model should leave its peak ratio essentially unchanged. If all three hold, mean-seeking is established as a distinct, predictable, and unsafe failure mode of neural stress reconstruction, with a cheap fix. If the mixture network or matched-cost diffusion equals the categorical head, the diagnostic result stands and the method claim narrows accordingly.