AI Meets Grease: Machine Learning Predicts How Graphene Keeps Machines Cool and Wear-Free

Grease is the unsung hero of the mechanical world. It sits quietly inside rolling bearings, sealed gearboxes, and countless maintenance-limited machines, doing the unglamorous work of keeping metal surfaces from grinding each other into oblivion. Yet for all its ubiquity, grease remains surprisingly difficult to understand. Unlike oil, which flows freely and continuously replenishes the surfaces it protects, grease is a complex semi-solid whose behavior changes dramatically over time. Now, a team of researchers has unveiled a hybrid machine learning framework that can predict, minute by minute, how graphene-enhanced grease will heat up, wear down, and transition between lubrication regimes—a breakthrough that could reshape how engineers design lubricants for everything from electric vehicles to wind turbines.

The study, published in the journal Results in Engineering by Ethan Stefan-Henningsen and Amirkianoosh Kiani, tackles a fundamental problem in tribology, the science of friction, wear, and lubrication. When a grease-lubricated machine starts operating, the thickener network that gives grease its structure is mechanically worked, gradually releasing base oil into the contact zone. This creates a transient, time-dependent lubrication state: the system begins in a high-drag churning phase and eventually settles into a bleeding phase, where lubrication depends on the released oil and the thermal state of the contact. As the supply of bleeding oil becomes insufficient, the lubricant film thins and the contact shifts from full-film elastohydrodynamic lubrication toward mixed or boundary regimes, where surfaces touch more directly and wear accelerates. Starvation, the researchers emphasize, is not an exceptional condition but the dominant operating state in many grease-lubricated rolling contacts.

What makes the new framework distinctive is its refusal to treat these phenomena as separate prediction problems. Most previous machine learning applications in tribology have focused on isolated outputs—a friction coefficient here, a wear rate there, a contact temperature somewhere else. The Canadian team instead built an integrated pipeline in which experimentally anchored temperature histories serve as a shared state variable linking everything else. The framework begins with nonparametric temperature prediction using inverse distance weighting, a transparent interpolation method chosen deliberately over flashier alternatives like neural networks or Gaussian processes because the dataset was small and the researchers wanted to avoid overfitting while exactly recovering measured trajectories at known experimental points. Compact thermal descriptors extracted from these predicted trajectories—final temperature, maximum temperature, mean temperature, cumulative thermal exposure, and early heating rate—then feed into ridge-regularized regression models that estimate wear scar area, wear depth, and coefficient of friction.

The experimental foundation of the work was a carefully controlled block-on-ring sliding configuration, in which a stationary aluminum block was pressed against a rotating hardened steel ring at a constant speed of 1000 rpm under a 4.9-newton load. The team formulated greases containing 0, 0.25, 0.5, 0.75, 1, and 2 percent graphene by weight, dispersing the nanoplatelets—supplied with lateral dimensions ranging from roughly 50 nanometers to 5 micrometers and flake thickness below 1.7 nanometers—in a mineral oil before high-shear mixing into a commercial lithium grease matrix. Each formulation was tested ten times, with an infrared thermal camera capturing thirteen-point temperature trajectories at 30-second intervals over 360-second runs. Standardized ASTM four-ball wear tests and apparent viscosity measurements by an independent laboratory rounded out the characterization.

The results are striking. The control grease, with no graphene at all, reached a peak temperature approaching 90 degrees Celsius by the end of testing. The 0.5 percent graphene formulation, by contrast, peaked at roughly 61 to 62 degrees—a reduction of about 30 percent. The mechanism is twofold: graphene’s lamellar structure reduces interfacial shear, so less frictional work is converted into heat, and the nanoplatelets’ exceptional in-plane thermal conductivity helps spread heat through the contact region. Equally important, the benefit was not monotonic. At higher concentrations of 1 and 2 percent, peak temperatures crept back up, and the coefficient of friction at 2 percent actually exceeded that of the base grease. The authors point to concentration-dependent dispersion effects—possible particle clustering or changes in the grease microstructure—as plausible, though not directly confirmed, explanations.

The wear predictions tell a similar story. The base grease produced a wear scar area of approximately 47.8 square millimeters, while graphene-containing greases between 0.25 and 0.75 percent produced scars of 30 to 32 square millimeters—a roughly 35 percent reduction. Wear depth dropped from about 87 micrometers for the control to 68 to 71 micrometers, with the minimum again near 0.5 percent. The coefficient of friction fell from approximately 0.092 for the control to about 0.087 at the optimum concentration. Validation metrics were impressive across the board: the framework achieved a coefficient of determination of 0.973 for temperature prediction, 0.955 for wear scar area, 0.954 for wear depth, and 0.921 for friction, with error distributions centered near zero and no significant systematic bias.

Perhaps the most scientifically rich output of the framework is its ability to reconstruct the invisible evolution of the lubrication state itself. Using Hertzian contact mechanics, the Hamrock-Dowson elastohydrodynamic film thickness correlation, temperature-dependent viscosity, and a phenomenological starvation model, the framework tracked how the effective film thickness collapsed by more than an order of magnitude during the 360-second test—from around 5 micrometers at the start to roughly 0.2 micrometers near the end. The Tallian lambda ratio, the dimensionless yardstick that compares film thickness to composite surface roughness, fell steadily from above 4, indicating full-film lubrication, to below 1, signaling a transition into the mixed regime where asperity interactions begin to dominate. Throughout, the Reynolds number remained on the order of 10 to the minus 3, confirming that viscous forces utterly dominated the thin lubricant film. Independent analytical verification showed all key lubrication quantities agreed with hand calculations within about 3 percent.

Herein lies the elegant explanation for why graphene-containing greases keep performing even as the hydrodynamic film decays. Once the contact transitions into mixed lubrication, direct asperity contact would normally drive friction and wear sharply upward. But graphene nanoplatelets form protective tribofilms and low-shear sliding layers at the interface, cushioning those asperity interactions. The framework’s energy analysis reinforces this picture: frictional power loss dropped from about 0.80 watts for the control to roughly 0.76 watts at the optimum concentration, and heat flux followed the same concentration-dependent trend. The lowest specific wear rate occurred between 0.25 and 0.5 percent graphene, tying improved thermal control directly to enhanced wear resistance. The authors are careful to note that the infrared measurements capture the spatially averaged surface temperature, not asperity-scale flash temperatures, though reduced interfacial heat generation makes lower flash temperatures physically plausible.

Compared with recent AI-assisted tribology studies—which have achieved near-perfect predictions for dry sliding of coated steels, magnesium composites, and self-lubricating materials—the present framework offers comparable accuracy while extending prediction into territory those studies never touched: grease-specific starvation-driven film decay, temperature-dependent viscosity reduction, lambda-ratio evolution, and lubrication regime transitions. This matters enormously for emerging applications. In electric drivetrains, where bearings run at high speeds under duty-cycled conditions, friction torque and temperature histories are increasingly used as practical indicators of lubrication quality and energy efficiency. A tool that can predict the entire transient timeline—not just an endpoint—gives formulators a way to design lubricants for the journey, not merely the destination.

The researchers are candid about the framework’s limits. Its generalization is constrained by the investigated graphene concentration range, the single grease chemistry, the modest dataset size, and the controlled block-on-ring conditions. Future work, they write, should proceed in stages: expanded testing across wider loads, speeds, temperatures, and durations; then different grease chemistries and additive systems; and finally component-scale and field-relevant validation. Still, the conceptual advance is clear. By fusing simple, transparent machine learning with classical lubrication physics, the study demonstrates that predictive tribology does not require opaque black boxes—it requires the right architecture, one in which temperature becomes the thread that stitches friction, wear, film formation, and energy dissipation into a single, physically interpretable story. For machines that quietly keep the modern world turning, that story may soon be written before the first revolution even begins.

Subject of Research: Hybrid machine learning prediction of thermal response, wear, and lubrication regime evolution in graphene-enhanced grease

Article Title: A hybrid machine learning framework for predicting thermal response, wear and lubrication regime evolution in graphene-enhanced grease

Article References: Stefan-Henningsen, E., & Kiani, A. (2026). A hybrid machine learning framework for predicting thermal response, wear and lubrication regime evolution in graphene-enhanced grease. Results in Engineering, 32, Article 113423. https://doi.org/10.1016/j.rineng.2026.113423

Image Credits: AI Generated

DOI: 10.1016/j.rineng.2026.113423

Keywords: graphene, grease lubrication, machine learning, tribology, wear prediction, elastohydrodynamic lubrication, starvation, friction, thermal response, lambda ratio, nanoparticle additives, predictive modeling