AI Learns to Predict Hidden Damage in Fretting Contacts in Under a Millisecond

Deep inside jet engines, hip implants, bolted joints, and electrical connectors, a quiet destroyer is at work. Fretting, the damage caused by tiny oscillatory movements between surfaces pressed together, can grind away metal, seed fatigue cracks, and quietly shorten the life of critical components. Now, a team of researchers has built a hybrid framework that pairs the physical rigor of finite element simulation with the speed of machine learning, allowing engineers to predict how much energy a fretting contact will dissipate and which slip regime it will fall into, in less than a millisecond per prediction. The work, published in Results in Engineering, could transform how industries screen materials and design interfaces against one of tribology’s most stubborn failure modes.

The challenge with fretting is that its behavior is fiercely nonlinear. The severity of damage depends on the normal load pressing the surfaces together, the displacement amplitude of the oscillation, the contact geometry, and the coefficient of friction. Even small changes in these parameters can shift the contact response from one regime to another, completely altering the dominant damage mechanism. When the contact sticks with only limited microslip at the edges, the system is in the partial slip regime, marked by low dissipated energy but high localized stress concentrations near the stick-slip boundary. When sliding occurs across most of the interface, the gross slip regime takes over, bringing higher energy dissipation and wear-dominated damage. Between these extremes lies the mixed slip regime, a transitional condition where stick and slip coexist, and where misclassification can lead engineers to draw the wrong conclusions about whether fretting wear or fretting fatigue will govern component failure.

Traditionally, understanding this behavior has demanded either exhaustive experimental campaigns or high-fidelity finite element simulations, both of which are expensive and slow. Each simulation in the new study, run in Abaqus on a ball-on-flat contact configuration, took between two and six hours on a workstation, and generating the complete dataset of 170 simulations consumed roughly 800 hours of computation. Experiments are no faster, since capturing the nonlinear dependence of contact variables and regime transitions requires a large number of carefully controlled tests. The research team, led by Avneesh Kumar with Marek Vostrak and Sarka Houdkova, recognized that if a machine learning model could learn the mapping from operating conditions to fretting response, the costly simulations would only need to be run once to build a training dataset.

Before any machine learning could begin, however, the team had to prove that their simulations were physically trustworthy. They verified the normal contact response against the classical Hertzian theory of elastic contact, comparing predicted contact radii and maximum contact pressures with analytical solutions across multiple loading conditions. Mesh sensitivity analyses showed that refining the contact mesh from 20 to 10 micrometers changed the predicted dissipated energy by only about 1.28 percent, confirming numerical convergence. The tangential behavior was verified against the Cattaneo-Mindlin theory of partial slip: at a ratio of maximum tangential force to friction capacity of 0.25, the simulated and analytical stick-zone radii differed by just 0.43 percent, and at 0.50 the difference was 1.99 percent. Finally, the team compared a simulated hysteresis loop against an experimental one from an alumina ball pressed against a cold-sprayed IN625 flat specimen under a 5-newton load and 10-micrometer displacement amplitude. The dissipated energies agreed within 5.8 percent, providing experimental confidence in the model’s cyclic predictions.

With a verified model in hand, the researchers generated their dataset by systematically varying normal load, displacement amplitude, and coefficient of friction. From each simulation they extracted the tangential force-displacement hysteresis loop, whose enclosed area yields the dissipated energy per cycle, an energy-based damage indicator rooted in the influential work of Fouvry and colleagues. Slip regimes were labeled using three simultaneous criteria: the normalized tangential force ratio, the ratio of dissipated energy to total input energy, and the characteristic shape of the hysteresis loop, whether linear, elliptical, or parallelogram-like. Cases where the indicators disagreed were excluded, ensuring clean labels. Crucially, the team also engineered contact-mechanics-guided features, such as the nominal friction capacity, a nominal frictional work proxy, the Hertzian contact radius, and the normalized displacement ratio, all derived deterministically from the input conditions rather than from simulation outputs, to avoid leaking target information into the predictors.

Four supervised learning algorithms were then trained for both regression and classification tasks: linear and logistic regression as interpretable baselines, support vector machines, random forests, and extreme gradient boosting, or XGBoost. Hyperparameters were tuned with repeated five-fold cross-validation, and a held-out test set was reserved for final evaluation. For dissipated energy prediction, XGBoost regression with the derived features delivered the strongest held-out performance, achieving a coefficient of determination of 0.966 and a root mean square error of 0.510, along with the highest mean cross-validated R-squared of 0.912. Support vector regression achieved the lowest average cross-validated error, though with greater variability across folds. No single model dominated every metric, a finding the authors report with appropriate caution.

The derived features proved decisive, and in a model-dependent way. Linear regression, nearly useless with raw inputs at a test R-squared of 0.471, jumped to 0.900 once the physics-informed features were added, indicating that the transformations made the underlying relationship compatible with a linear formulation. Random forests and XGBoost also improved their cross-validated consistency, while the support vector machine performed comparably with either feature set. For classification, logistic regression with derived features correctly identified every sample in the held-out test subset, while the baseline support vector classifier achieved the highest mean cross-validated accuracy at 0.949. The mixed regime proved hardest to classify, as expected for a transitional state sharing characteristics with both neighbors, and confusion matrices showed most errors involved mixed regime samples being labeled as partial or gross slip.

The team went beyond standard validation with blocked, leave-region-out tests, withholding entire ranges of a single operating parameter to probe whether models could extrapolate. Performance depended strongly on which parameter was withheld: predictions transferred well across unseen friction coefficient values, with R-squared as high as 0.954, but high normal load and high displacement regions proved harder to generalize. Interpretability analyses using permutation importance and SHAP values revealed that friction capacity and displacement-related features, particularly the normalized displacement relative to the Hertzian contact radius, were most strongly associated with regime predictions, though the authors emphasize these are model-specific associations rather than proof of physical causality, especially given the correlations among derived features.

The payoff is striking: once the initial dataset is generated and models trained, a prediction takes less than a millisecond, compared with hours for a new simulation. Hyperparameter optimization for all eight models took only about eleven minutes. This makes the framework a practical surrogate for preliminary screening and parametric assessment within the investigated ball-on-flat design space, with applications ranging from biomedical implant coatings to bolted aerospace joints. The authors are candid about limitations: the model assumes smooth surfaces, a constant friction coefficient, isothermal conditions, and no third-body debris effects, so transfer to other geometries and materials remains constrained. Future work targeting surface roughness, temperature effects, and more representative contact geometries will determine how far this simulation-plus-machine-learning paradigm can reach. But the core message stands: by anchoring data-driven models in verified contact mechanics, engineers can now interrogate fretting damage at the speed of thought rather than the speed of supercomputing.

Subject of Research: A hybrid finite element and machine learning framework for predicting dissipated energy and slip-regime classification in fretting contacts

Article Title: Finite-element and machine-learning framework for dissipated-energy prediction and slip-regime classification in fretting contacts

Article References: Kumar, A., Vostrak, M., & Houdkova, S. (2026). Finite-element and machine-learning framework for dissipated-energy prediction and slip-regime classification in fretting contacts. Results in Engineering, 32, Article 113296. https://doi.org/10.1016/j.rineng.2026.113296

Image Credits: AI Generated

DOI: 10.1016/j.rineng.2026.113296

Keywords: fretting, tribology, machine learning, finite element analysis, dissipated energy, slip regime, XGBoost, contact mechanics, hysteresis loop, wear prediction, fretting fatigue, SHAP analysis