Beyond visible rust: a proposed AI framework for assessing corroded bolted joints

From appearance to function

Bolted joints depend on more than the condition visible in a photograph. Their behavior also reflects friction, surface contact, tightening history, joint geometry and preload—the clamping force established during assembly. Corrosion can alter surface roughness, create pits and change contact conditions, affecting friction and the retention of that force.

These relationships matter when deciding whether returned products should be reused, refurbished, disassembled or recycled. Their exposure and maintenance histories may be unknown, making appearance alone an incomplete basis for assessing remaining function.

In a preprint posted to arXiv on November 6, 2026, Nehal Afifi and Sven Matthiesen of Karlsruhe Institute of Technology, together with Gerhard Kachel of Offenburg University of Applied Sciences, propose a workflow connecting inspection evidence to selected mechanical indicators. The contribution is a methodological architecture, not a demonstrated prediction system: the paper reports no experimental accuracy results or validated corrosion equations.

Bringing different views together

The proposed workflow starts with synchronized measurements. Camera images would record broad rust patterns and joint geometry; microscopy would capture local pits, oxide structures and surface damage. Roughness measurements, exposure information and joint specifications would provide additional context. Mechanical tests would supply the functional quantities the system is trained to predict.

Traceability is essential. Each microscopic patch must be associated with a specimen, inspection time, surface location and corresponding mechanical record. For equation discovery, the authors specify repeated or re-locatable observations of the same regions, multiple specimens per exposure/preload condition, and a baseline followed by at least three post-exposure intervals.

A learning model would align corresponding camera and microscope patches in a shared numerical representation. It would then combine those representations with roughness, exposure and joint information. This step is intended to bridge scales, rather than directly turn a photograph into a mechanical verdict.

A calibrated map, not a chemical measurement

The next proposed output is a surface map of corrosion severity, with values between zero and one. It would combine evidence such as rust coverage, pit distribution, surface-height changes and material loss when available.

The authors emphasize that this map is a calibrated surrogate: a constructed representation tied to observable degradation, not a directly measured electrochemical state. Missing measurements would be omitted from calibration rather than invented as ground truth.

Mechanical prediction targets would remain separate from this calibration. That separation is intended to prevent the corrosion map from simply encoding the response it is later supposed to help predict.

Searching for useful equations

With severity maps collected over time, the framework would search for partial differential equations describing how those maps evolve. Such equations express changes across space and time.

A genetic algorithm would explore combinations of candidate mathematical terms, penalizing excessive complexity and physically implausible behavior. For noisy reconstructed maps, the authors favor a formulation that averages discrepancies over regions and time intervals, reducing reliance on noisy point-by-point derivatives.

Any resulting equation would describe an effective pattern of degradation at the modeled scale—not establish a complete chemical mechanism. Equations containing learned, otherwise uninterpreted features would be treated as predictive tools rather than mechanistic explanations.

A physics-informed neural network would use the selected equation as a constraint. A separate functional model would predict friction coefficient from corrosion descriptors and joint information. Preload loss would be an additional target when clamp-force data are available. Predicting complete torque-angle or force-displacement curves is explicitly left for later work.

What must be shown

The validation plan asks whether this elaborate pipeline improves on simpler alternatives. Tests would compare it with image-only and scalar-input models, as well as a neural predictor without the equation constraint.

Entire specimens—not individual image patches—would be separated between training and evaluation to avoid information leakage. Forecasting would also be tested on exposure intervals excluded from fitting and model selection. Synthetic fields with known equations would precede interpretation of experimental discoveries.

The authors identify clear failure conditions: severity maps that disagree with independent damage measurements, equations that change under plausible measurement perturbations, or improvements confined to training data. Misaligned images can create artificial spatial gradients, while poorly balanced constraints can destabilize neural-network training.

The potential application is better-informed inspection, including recognizing when uncertainty warrants further examination. Whether the framework actually improves functional predictions and supports reuse decisions remains an empirical question for the planned experiments.