Steel in concrete usually corrodes in small spots. Chloride salts from seawater or de-icing seep through the concrete cover and break down the thin protective film on the rebar at isolated points. The result is a pit that eats steel quickly and shows almost nothing at the surface. Inspectors can measure electrical potentials on the concrete surface. Recovering a pit's position, size, and corrosion rate from those measurements is an inverse problem: given the surface readings, work back to the hidden cause. The most careful current approach, electrochemical tomography, handles this with one large stochastic solver that fits the whole slab at once. That solver inherits two failures that the machine-learning error-correction literature knows well. It spends effort everywhere, although most of a bridge deck is passive steel. It also over-corrects: for small anodes it produces physically implausible spikes in current density, which the authors call the "anode size paradox."
We propose to apply the design that fixed the same two failures in text and image correction. A cheap, learned detector first flags the few regions of a potential map that plausibly contain an active anode. An expensive, physically bounded finite-element inversion then runs only inside those flagged regions. It also runs only on the parameters the data can actually determine there. The claim to test is specific. At structure scale, gating plus domain restriction should match or beat whole-slab inversion on pit localization and anodic-current error, cut false pits on passive steel, and use a small fraction of the compute. We will separately measure how much of any gain comes from the gating and how much comes from simply imposing physical bounds, because the second could explain everything.
What the two directions contribute
The error-correction side supplies the method. Chen et al. split grammatical error correction into a token tagger and a corrector that rewrites only the flagged spans. On BEA-2019 this reduced decoder steps from 21,065 to 7,647 and still improved F0.5 from 65.5 to 66.9. AIP-Tohoku added a sentence-level gatekeeper that lets clean sentences bypass the corrector, which alone raised precision by over 3 points. The closest structural analogue is Zung et al. in connectomics. There, a 3D convolutional detector flags split and merge errors, and the union of flags forms an "advice mask" that confines an expensive corrector to about 10% of a volume of more than 2.3 billion voxels. Errors fell from 944 to 305. We reuse the advice-mask idea directly: the detector's output becomes the spatial domain of the inversion.
These papers also constrain the design. Each reports the same cascading-error weakness: whatever the detector misses, the corrector never sees. Zung et al. further show that the corrector cannot fix errors below its working granularity. Our evaluation must therefore report missed anodes as a first-class metric and not bury them in an average.
The corrosion side supplies the problem and the physics. The tomography paper by van Ede et al. is the method we build on and the baseline we must beat. It couples small perturbation currents, a 3D finite-element model of electrical conduction, and stochastic inversion for anode location, anode width, and exchange current density. It also measures information gain with Kullback–Leibler divergence, which quantifies how much the data narrows the estimate compared with the prior. Its findings motivate our proposal in two ways. Information collapses when cover exceeds about 5 cm, and small anodes produce inflated current densities. Both suggest that anode width and current density cannot be told apart for small, deep anodes. Asking one solver to return both everywhere invites artifacts.
The corrosion meta-analysis of Coelho et al. finds rebar corrosion to be the most scattered, least stable domain for machine-learning regressors, particularly for localized pitting. This result constrains us. It argues against replacing physics with a single end-to-end network, and it is why our learned component only localizes, leaving quantification to the physical solver.
The chloride-transport papers set the scenario distribution. Wang et al. show that carbonation decomposes Friedel's salt and releases bound chloride ahead of the carbonation front. The stress-diffusion study shows that chloride ingress is non-monotonic in sustained compressive stress, slowing at 0.3 f_c and speeding up at 0.5 f_c once microcracking begins. We use these mechanisms to decide where synthetic anodes appear in training and test slabs. The goal is that the detector is exercised on realistic spatial patterns and not on uniformly scattered pits.
Core Claim
We claim that separating "is there an anode here, and roughly where" from "what are its size and kinetics" changes what the inversion is asked to do. It does not just make the inversion faster. The detector's output has two parts: a region mask, and a flag saying whether the local cover and resistivity leave anode width identifiable. The inversion then solves a smaller, better-posed problem inside the mask. Where width is identifiable, it estimates width and exchange current density under Butler–Volmer bounds, the standard electrode-kinetics equation that limits physically plausible current densities. Where width is not identifiable, it reports only total anodic current, the product the surface data can actually determine, together with an honest width posterior.
The novelty lies in the combination. Neither the tomography work nor the connectomics work gates both the spatial domain and the parameter set of a physics-based inversion using a learned detector. No work in the file does this for corrosion.
Research Program
The first component is a synthetic generator. It extends the van Ede finite-element setup from single-anode slabs to deck-scale panels of several square meters. Anode placement is drawn from chloride-exposure scenarios: thin-cover zones, cracked zones under high sustained stress, and carbonated zones with chloride release. Cover ranges from under to over 5 cm, and resistivity spans values below and above 100 Ω·m. Each panel yields surface potential maps under several perturbation currents, along with ground-truth anode geometry and kinetics.
The second component is the detector. It is a small 2D convolutional network on the stacked potential maps. It outputs a per-location anode probability and a per-region identifiability score. The identifiability score is trained to predict the Kullback–Leibler information gain that the full inversion would achieve, a target the generator can compute offline.
The third component is the corrector. It is the van Ede stochastic inversion, unchanged in physics, restricted to the dilated advice mask. Its parameter set is switched by the identifiability flag. A recall-oriented detector threshold, together with dilation of the mask, is our main defense against cascading misses. We will sweep both settings to trace the trade-off between compute and missed anodes.
Baselines
The naive reference is standard half-cell potential mapping with a fixed potential threshold. It is what practitioners use, and it indicates corrosion probability but gives no size or rate. It runs on the same synthetic maps, and localization is scored against the same ground truth. Any method that cannot beat it on localization F1 is not worth its cost.
The strongest existing method is monolithic stochastic inversion as in van Ede et al., run over the whole panel. We found no reported localization error or anodic-current error on Lacuna for that method; its information-gain values appear only graphically. We will therefore reimplement it on our generator with the same finite-element mesh, forward model, sampler, and number of forward solves as our pipeline's total budget, plus a second run with unlimited budget. Beating it means lower median error in anodic current and fewer spurious anodes on passive panels at equal compute. It also means comparable accuracy at no more than a quarter of its compute, with anode recall no worse than two points below it.
The ablation that isolates our key ingredient is monolithic inversion with the same Butler–Volmer bounds and priors we give the corrector, but no detector and no mask. If this matches our accuracy, the gain comes from bounds and not from decoupling, and our remaining contribution is compute alone. A second ablation keeps the mask but removes the identifiability switch, which isolates the parameter gating. A third baseline is an end-to-end neural network that regresses anode parameters directly from the maps, given the same training data and tuning budget as our detector.
Why not just…?
Why not just bound the current density in the existing solver? This is the most serious threat. It may remove the anode-size spikes without any detector. The bounded-monolithic ablation tests exactly this. We expect bounds alone to trade spikes for width estimates dominated by the prior, while still spending compute on passive steel and still producing occasional false anodes there. If the ablation shows equal accuracy and false-alarm rate, we will report the method as a compute saving and not as an accuracy gain.
Why not just report the full Bayesian posterior and let wide marginals signal non-identifiability? Over a small slab this is principled, and van Ede's framework already supports it. At deck scale, however, a joint posterior over many candidate anodes is expensive to sample and hard to read. The identifiability flag is a cheap, learned stand-in for the information gain the posterior would show. We will check how well the flag agrees with posterior-derived information gain on held-out panels. If agreement is poor, the flag is not earning its place.
Why not just train one network to predict everything from the maps? Coelho et al. report that monolithic regressors are least reliable precisely for rebar pitting. The end-to-end baseline will show whether that holds on our clean synthetic data. If it does well in-distribution, the decisive check becomes out-of-distribution panels, with cover and resistivity outside the training range. There the physics-based corrector should degrade more gracefully.
Proposed Experiments
The main experiment compares the six systems on held-out synthetic panels, grouped by cover (under versus over 5 cm) and resistivity (under versus over 100 Ω·m). The metrics are anode-localization F1, median and 90th-percentile relative error in total anodic current, width error on panels flagged identifiable, spurious anodes per square meter on fully passive panels, and finite-element forward solves per panel.
The first decisive check is the bounded-monolithic comparison described above. The second is a sweep of detector threshold and mask dilation, which traces recall against compute and shows whether cascading misses can be held below the monolithic solver's own miss rate. The third is a scenario-shift test: we train anode placement on chloride-only scenarios and test on carbonation-with-chloride-release and high-stress cracking scenarios. This tests whether a detector trained on one exposure pattern still finds anodes under the chloride-transport mechanisms of Wang et al. and the stress study.
As a cross-domain calibration, the connectomics pipeline confined its corrector to about 10% of the volume. If our masks cover a similar fraction of deck area at matched recall, the compute argument is established. The file contains no physical tomography measurements. A small laboratory validation on slabs with induced chloride anodes would be the natural next step, but it lies outside the resources listed here.
References
- Inverse modelling of potential measurements on reinforced concrete to assess localized corrosion: Sensitivity with respect to the corrosion state and concrete properties
- Improving the Efficiency of Grammatical Error Correction with Erroneous Span Detection and Correction
- The AIP-Tohoku System at the BEA-2019 Shared Task
- An Error Detection and Correction Framework for Connectomics
- Reviewing machine learning of corrosion prediction in a data-oriented perspective
- Research on the Influence of Carbonation on the Content and State of Chloride Ions and the Following Corrosion Resistance of Steel Bars in Cement Paste
- Chloride Transport Modeling of Binary Mineral Admixture High-Performance Concrete Under Sustained Compressive Stress