TL;DR
This paper introduces a physics-informed multimodal learning framework that bridges the gap between visual corrosion inspection and the actual mechanical performance of bolted joints. By combining microscopic imaging, PDE discovery, and Physics-Informed Neural Networks (PINNs), the framework allows engineers to predict critical functional indicators like friction and preload loss, providing a scientific basis for deciding whether to reuse, refurbish, or recycle components in a circular economy.
What Problem Does This Paper Address?
In the push toward a circular economy, "circular factories" need to decide if returned components—like the ubiquitous bolted joint—are still fit for purpose. However, assessing a corroded bolt is notoriously difficult. Corrosion doesn't just look bad; it fundamentally changes surface morphology, roughness, and contact mechanics, which in turn alters the joint's friction coefficient and its ability to retain preload.
Currently, engineers are caught between two extremes: complex electrochemical models that require state variables impossible to measure in the field, and "black-box" computer vision models that can spot rust but can't tell you if the bolt will fail under load. This paper addresses the missing link: a structured, physically constrained path that connects feasible inspection data (like smartphone photos and microscopic patches) to reliable functional predictions.
Methods
The authors propose a multi-stage framework that moves from raw pixels to physical laws:
- Multimodal Similarity Learning: Because microscopic images and standard camera photos capture the same degradation at different scales, the framework uses a two-branch "Siamese" encoder to align these observations into a shared latent space.
- Corrosion State Inference: These aligned features are fused with roughness measurements and joint descriptors to construct a calibrated corrosion-severity field ($c$). This field acts as a surrogate for the physical state of the surface.
- PDE Discovery via Genetic Algorithms: Instead of assuming a physical law, the framework searches a library of candidate mathematical terms (like diffusion or transport) using a genetic algorithm to "discover" the effective mesoscale partial differential equation (PDE) that governs how the corrosion field evolves over time.
- Physics-Informed Neural Networks (PINNs): Finally, a PINN is trained to predict functional targets (friction coefficient $\mu$ and preload loss $\Delta F_p$). Crucially, the network is constrained by the discovered PDE, ensuring its predictions remain physically plausible.
Datasets
The paper outlines a rigorous Design of Experiments (DoE) rather than using a static, pre-existing dataset. The methodology requires:
- Synchronized Multimodal Records: Paired camera images, microscopic patches, and roughness profiles.
- Functional Response Data: Torque-angle and torque/clamp-force curves (per ISO 16047).
- Time-Series Observations: At least three post-exposure intervals (using accelerated corrosion tests like ISO 9227) to allow for PDE discovery.
- Geometrical Variety: Different bolt sizes, strength grades, and preload levels to ensure the model generalizes across joint configurations.
Key Insights and Ideas
- Images as Evidence, Not Labels: A core shift in this work is treating images not as final classification targets, but as spatial evidence for a continuous degradation field. This prevents "visual texture" from being misinterpreted as physics without proper calibration.
- Discovering "Effective" Laws: The authors acknowledge that full electrochemical modeling is often impossible for returned products. By using PDE discovery, they find "effective mesoscale laws" that are scientifically bounded and useful for prediction, even if they don't capture every atomic interaction.
- Latent-Field Interface: By creating a "latent-field interface" (shown in Figure 2), the model can ingest heterogeneous data—from a low-res phone photo to a high-res microscopic pit measurement—and map them onto the same physical surface frame.
Key Results
As a framework proposal, the paper focuses on the mathematical and architectural formulation:
- Traceability: The framework provides a clear "decision chain" where every functional prediction is linked back to specific surface observations and discovered physical constraints.
- Falsifiability: The authors include a validation strategy where the framework is considered "falsified" if the discovered PDE terms are unstable or if the image-derived fields fail to improve prediction over simple scalar descriptors.
- Uncertainty Quantification: The inclusion of an "Uncertainty Head" in the PINN architecture (see Figure 3/4) allows the model to signal when it lacks sufficient data to make a safe reuse recommendation.
Limitations and Future Work
The authors identify several challenges:
- Data Intensity: The need for repeated, re-localizable observations of the same surface regions over time is a high bar for data collection.
- Registration Sensitivity: Small errors in aligning microscopic patches to camera images can introduce "artificial gradients" that confuse the PDE discovery algorithm.
- Future Implementation: The next step for this research is the implementation of the framework on a dedicated experimental dataset with specimen-level splits to verify the "central empirical question": do these image-derived fields actually improve held-out predictions of joint function?
Summary updated