Machine Learning Cracks the Code of Nitinol Wear, a Metal That Remembers Its Shape

Nitinol, the remarkable nickel-titanium alloy that can remember its shape and flex thousands of times without deforming, has long been prized for everything from medical stents to aerospace couplings and dent-resistant bearings. Yet one stubborn question has haunted engineers who deploy this superelastic metal: how fast will it wear out under real sliding contact? Predicting wear in Nitinol has resisted the simple equations that work for ordinary steels, because the alloy’s response shifts dramatically with heat treatment, load, speed, and test duration. Now a new study published in the Journal of Materials Science shows that machine learning, applied with unusual rigor, can finally pin down those elusive wear and friction trends, and even recover a generalized wear law that the classical textbook equation could not.

The research, carried out by Samuel Onimpa Alfred of the Department of Aerospace Engineering at the University of Michigan, belongs to a rapidly growing field called triboinformatics, where artificial intelligence is used to decode the messy, nonlinear world of friction, lubrication, and wear. Tribology, the science of interacting surfaces in relative motion, is not a niche pursuit. Earlier studies cited in the work estimate that friction and wear-related losses consume a staggering share of global energy, contributing substantially to costs and emissions across industry. Making surfaces last longer, in artificial hip joints or in jet engine bearings, translates directly into energy saved and devices that survive longer inside the human body or inside an aircraft engine.

What makes Nitinol such a fascinating test case is its shape-memory and superelastic character. Deform it and it snaps back; heat it past a critical temperature and it returns to a previously memorized configuration. But these same properties make its tribological behavior notoriously complex. In this study, Alfred examined four distinct Nitinol conditions: an equiatomic titanium-nickel alloy, and three versions of a nickel-rich composition known as 60NiTi that had been aged, annealed, or solution-treated. Each heat treatment changes the alloy’s microstructure, hardness, and elasticity, and therefore changes how it wears when scraped against a counterface in dry, reciprocating sliding. Classical wear laws, which tend to assume wear scales simply with load and sliding distance while inversely scaling with hardness, capture these shifts poorly.

To tame that complexity, Alfred assembled a dataset of 336 individual measurements drawn from previously published, peer-reviewed reciprocating dry-sliding experiments, covering weight loss, cumulative wear over time, and steady-state coefficient of friction for all four material conditions. The modeling strategy then subjected a battery of algorithms to a demanding examination. Instead of splitting the data randomly, the study used grouped cross-validation, meaning that entire series of wear tests were withheld from the model during training. This is the scientific equivalent of asking a student to answer questions about chapters of a book they were never allowed to read, and it guards against the overly optimistic predictions that plague many machine-learning studies in materials science.

Even under this harsh test, gradient boosting, an ensemble method that builds a predictive model from many sequentially corrected decision trees, emerged as the clear winner. It predicted weight loss with an R-squared of 0.92 and the coefficient of friction with an R-squared of 0.91, beating support-vector regression, random forests, and simple linear baselines. When the cross-validation was made less restrictive and random splits were allowed, those scores climbed to 0.98 for wear and 0.94 for friction. The result confirms that gradient boosting does not just memorize data; it learns genuinely transferable relationships between operating conditions and tribological outcomes, even for material-test combinations it has never encountered.

Crucially, the study did not stop at prediction. Using SHapley Additive exPlanations, or SHAP, a technique borrowed from cooperative game theory that assigns each input variable its fair share of credit for a model’s output, Alfred opened the black box. The analysis revealed that applied load is the dominant driver of wear, while oscillation frequency dominates friction behavior, with higher frequencies associated with lower friction coefficients, a trend consistent with frictional heating at the sliding interface. This kind of transparency matters enormously for engineers, because a model that merely outputs numbers without explanations offers no guidance on which design levers to pull.

Perhaps the most striking achievement is a deep learning result: a gated recurrent unit network, a type of neural network designed for sequential data, reproduced the full time-resolved wear trajectories of completely unseen tests with an R-squared of 0.91. In other words, given the early portion of a wear test, the network could accurately trace how material loss would accumulate over the entire remaining test, test after test, across all four heat treatments. That capability opens the door to digital wear forecasting, where a short initial experiment or monitoring window could stand in for long and expensive laboratory campaigns.

The study then confronted the granddaddy of wear science, the Archard equation, formulated in 1953, which states that wear volume is proportional to load and sliding distance and inversely proportional to hardness. When that classical law was fitted to the Nitinol dataset, it managed an R-squared of only 0.27, and performed even worse, going negative, when hardness was imposed rather than fitted. Alfred instead let the data speak, recovering a generalized Archard-type law in which the exponents on load, sliding distance, frequency, and hardness are free parameters. The resulting equation, weight loss equals 0.0425 times load to the power 0.78, sliding distance to the power 0.54, frequency to the power minus 0.14, and hardness to the power minus 0.26, described all four heat treatments with an R-squared of 0.79, using only measured hardness rather than material identity labels.

The exponents themselves tell a physical story. The sub-linear load exponent of 0.78 suggests that superelastic Nitinol distributes contact stress in a way that softens the wear increase as loads climb, while the negative frequency exponent quantifies the frictional-heating effect seen in the SHAP analysis. Most intriguingly, the fitted wear coefficient for each material correlated almost perfectly, at r equals 0.97, with the ratio of elastic modulus to hardness, a dimensionless quantity long championed in surface engineering as an indicator of elastic, wear-tolerant contact. When the modulus-to-hardness ratio served as the sole material descriptor in the generalized law, the fit reached an R-squared of 0.83, essentially matching models that knew which alloy they were dealing with.

For a metal that must survive inside arteries, bearings, and aerospace mechanisms without the benefit of lubrication, these findings provide something genuinely new: accurate, transparent, and physically consistent models that connect processing, properties, and performance. A designer can now estimate how a given heat treatment, hardness, and duty cycle will translate into wear and friction, before a single prototype is machined. More broadly, the work is a template for how triboinformatics should be done, with strict grouped validation, interpretable explanations, and laws recovered from data that honor the physics of contact rather than discarding it. As Nitinol finds its way into ever more demanding applications, the machines that predict its wear are, fittingly, learning from the metal that never forgets.

Beneath the headline results lies a dataset with an unusually clean provenance. The 336 measurements were not generated afresh for the modeling study but were compiled from two previously published, peer-reviewed experimental campaigns on superelastic TiNi and 60NiTi, with every table—loads, frequencies, durations, sliding distances, specimen masses before and after testing, cumulative weight loss, and steady-state friction coefficients—reproduced in the new paper’s appendix. That decision to expose the full experimental record alongside the models is itself a small contribution to a field where data scarcity and fragmentation remain the chief obstacles to progress.

The geometry underlying those tables also rewards a closer look. Each test used a 5.03 millimeter reciprocating stroke, so every cycle covered just over a centimeter of sliding, and the sliding distances reported for all four material conditions satisfy an exact arithmetic relation linking distance to frequency and test duration. All fifty-six wear-time series were strictly monotonic, with no missing entries, meaning the recurrent network tasked with reconstructing wear trajectories never had to impute gaps—an often unappreciated advantage when deep learning meets sparse laboratory data.

The strong correlation between the fitted wear coefficient and the elastic-modulus-to-hardness ratio also has a pedigree worth noting. Surface engineers have argued for decades that this ratio, rather than hardness alone, governs how well a material tolerates elastic contact and resists abrasion, particularly for coatings and for alloys whose elastic resilience absorbs deformation that would otherwise be permanent. The Nitinol results give that long-standing heuristic a quantitative, data-driven endorsement for shape-memory metals specifically.

The study situates itself in a broader movement. Recent systematic reviews of machine learning in tribology have catalogued a wave of applications, from aluminum-matrix composites to modified zinc alloys to diamond-like carbon coatings, where algorithms predict friction and wear from operating parameters. What distinguishes the present work within that wave is its insistence on withholding whole test series during validation and on recovering an interpretable wear law from the same data used to train opaque models. The author notes that the code underlying the analysis is available on reasonable request, and the supplementary data file, roughly 800 kilobytes in spreadsheet form, invites others to replicate or extend the models for their own shape-memory alloy systems.

Subject of Research: Machine learning prediction of wear and friction behavior in heat-treated Nitinol shape-memory alloys.

Article Title: Triboinformatic modeling of nitinol alloys under different heat-treatment regimes

Article References: Alfred, S. O. (2026). Triboinformatic modeling of nitinol alloys under different heat-treatment regimes. Journal of Materials Science. https://doi.org/10.1007/s10853-026-13713-9

Image Credits: AI Generated

DOI: 10.1007/s10853-026-13713-9

Keywords: Nitinol, tribology, machine learning, wear prediction, shape-memory alloys, gradient boosting, SHAP, Archard equation, deep learning, heat treatment, friction, triboinformatics