Predicting Gear Scuffing Onset from Contact Temperature History
Aerospace power reduction gearboxes run at the harsh end of the mechanical design envelope. Shaft speeds climb past twenty thousand revolutions per minute, lubricant films shrink to a few hundred nanometres, and meshing teeth carry loads that fluctuate with every passing turbine stage. Under those conditions, gear scuffing remains one of the most disruptive failure modes a designer can face. A localised weld forms between the flanks of meshing teeth, material transfers across the contact, and a surface that was mirror-polished a moment ago is suddenly scarred. The cost shows up in unscheduled engine removals, in lost dispatch availability, and in the slow erosion of safety margins as the fleet ages.
For several decades the standard tools for predicting scuffing have been bulk-flash-temperature models combined with empirically derived load-speed-temperature limits such as the integral temperature criterion or the FZG scuffing load stage. Those tools are still useful, but they tend to treat each meshing event as a single point in steady state. They average across the gear and across the operating cycle, and they leave a great deal of the physical story on the table. Real gears experience a thermal history: a succession of hot flashes, conduction pulses into the tooth body, and a gradual rise in bulk temperature as the lubricant carries heat away. It is that history, not a single number, that governs whether the protective film survives.
A regression model that takes contact temperature history as its input and returns a probability of scuffing initiation is an attempt to bridge the gap between physically rich but unwieldy thermal simulation and the operational need for a fast, defensible go/no-go answer. By collapsing the temporal signal into a small set of engineered features and fitting a regression to bench and rig data, the engineer can score a candidate gear set in seconds while still respecting the physics that drives the failure mode. The work sits inside the design-of-experiments and tolerance-analysis methodology used by the OPTIMIZE Project website and reflects the same instinct the broader research community has applied to bearing fatigue and rolling-contact wear.
The Australian aerospace and defence research community has skin in this game. Engineers at the Defence Science and Technology Group at Edinburgh in South Australia, designers at Boeing Defence Australia at Williamtown and Brisbane, and academic groups at RMIT and the University of Melbourne all wrestle with the same question: how do you prove that a new gearbox architecture will survive its full life without building a fleet of prototypes first? A regression model that listens to temperature history is one of the more promising ways to answer that question with limited test articles, and it fits naturally into the test-cell workflow already in use at the major Australian propulsion facilities.
The Physical Story Behind Scuffing Initiation
Scuffing begins where the elastohydrodynamic lubricant film breaks down. In a well-designed aerospace gear pair, the film thickness is on the order of a few hundred nanometres, the contact pressure runs into the gigapascal range, and the local flash temperature can exceed two hundred degrees Celsius. When the film collapses, asperities touch, the local temperature spikes again, and the softer of the two materials begins to adhere to its partner. The first transfer particle is small, often invisible, but it sets off a chain reaction: the damaged surface traps heat, the next meshing event arrives before the surface has cooled, and the failure propagates within a few hundred cycles.
The temperature history at the contact captures all of this. It reflects the bulk oil temperature, the gear speed, the load per tooth, the slide-to-roll ratio at the pitch point, and the condition of the surface. A sensor placed in the tooth root or in a nearby oil jet cannot see the flash temperature directly, but it can see the integrated thermal response, and with careful calibration that response can be back-translated into a contact-zone estimate. That is the premise the regression model is built on: not a measurement of scuffing itself, but a measurement of the thermal world in which scuffing either survives or ignites.
Why a Single Threshold Is Not Enough
Engineers have long used a single temperature threshold as a proxy for scuffing risk, and for many years that was a reasonable approximation. The human brain is good at picking out a single bright signal in a noisy trace, the same kind of pattern-spotting that draws some people to keno for Australian players at the local RSL on a Saturday arvo, but the scuffing problem rewards a more disciplined approach. A single number cannot tell the difference between a tooth that has been simmering at a moderately elevated temperature for an hour and a tooth that has just seen a sharp transient spike during a torque excursion. The first tooth may be perfectly safe; the second may already be on the path to scuffing. Temperature history separates the two cases, and a regression model can be trained to weight the relevant features accordingly.
A typical feature set used in the regression includes the peak flash temperature, the time spent above a critical threshold, the rate of change of temperature at the trailing edge of the meshing event, the cumulative thermal dose, and the bulk oil temperature at the moment of contact. The regression learns to combine them in a way that matches the test outcomes, and in practice the largest coefficients usually sit on the peak flash temperature and the time-above-threshold, with smaller but non-trivial contributions from the rate-of-change and the cumulative dose. The result is a score that is monotonically related to scuffing probability and that behaves sensibly when extrapolated slightly outside the training envelope.
Building the Model on Test Data
The training data come from a structured design of experiments run on a back-to-back gear rig. The team uses a fractional factorial plan that varies speed, torque, oil temperature, oil type, and surface finish across a wide operating window. Each test point is instrumented with thermocouples in the tooth root, in the housing, and in the oil supply, and the contact temperature is reconstructed through an inverse heat-transfer calculation calibrated against a handful of directly instrumented teeth. The scuffing outcome is recorded as either survival to a defined cycle count or initiation within that count, giving the regression a clean binary target.
Fitting is carried out in two stages. A regularised linear regression is run first on the raw engineered features to give a transparent baseline that designers can audit by hand, and then a gradient-boosted regression is run on the same features to capture the non-linear interactions the linear model misses. Both models are cross-validated against a held-out test matrix, and the residuals are examined for systematic patterns. A good model shows residuals scattered randomly around zero; a poor model shows clustering by operating regime, which usually points to a missing feature. The team has found that adding a feature describing the variability of the thermal signal during the run, rather than only its mean level, materially improves the fit.
Where the Regression Sits in the Wider Toolbox
The regression approach is not the only game in town, and it is worth being honest about where it sits. A full thermo-elastohydrodynamic finite-element simulation remains the gold standard for absolute prediction, but it is slow, sensitive to mesh quality, and difficult to validate against rig data without a great deal of effort. An empirical formula such as the Blok flash temperature criterion is fast and easy to apply, but it has no memory of the operating history. The regression on temperature history is the middle path: fast enough to run inside a Monte Carlo tolerance study, respectful of the thermal physics at a level the empirical formula does not reach, and without the deep numerical resources of a full simulation.
| Approach | Physical fidelity | Computational cost | Use of measured history | Fit for tolerance studies | Interpretability |
|---|---|---|---|---|---|
| Full thermo-EHL finite-element model | High | High | Limited | Low | Low |
| Empirical formula (Blok, integral temperature) | Medium | Low | None | Medium | High |
| Regression on contact temperature history | Medium-high | Low | Strong | High | Medium |
| Physics-informed neural network | High | Medium at training, low at run | Strong | High | Low |
The table makes the trade-off explicit. The regression on temperature history is the option that gives a designer the most useful combination of speed, fidelity, and historical sensitivity for routine design work, while the full simulation is reserved for the most demanding cases and for the validation of the regression itself.
Practical Implications for Engine Programs
For an engine program, the practical value of the model is that it can be embedded in a tolerance study. Manufacturing variation means that every gear in the fleet is slightly different, and the cumulative effect of those small differences can shift the predicted scuffing margin by a non-trivial amount. By running the regression across a Monte Carlo sample of manufactured gear sets, the design team can quantify the probability of scuffing at any point in the operating envelope and feed that probability into the safety assessment. The same workflow extends to condition-based maintenance: as a real gearbox accumulates temperature history, the model can score that history against the scuffing risk and flag the unit before the failure becomes catastrophic.
The approach also has implications for the test campaign itself. Instead of running every test point to destruction, the team can use the regression as a real-time monitor during the test, terminating the test when the model predicts a high probability of imminent scuffing. That saves rig time, limits the damage to expensive test hardware, and generates a richer dataset because every test ends at a meaningful boundary rather than an arbitrary cycle count. Several Australian test facilities, which are already configured for long-duration endurance runs on geared propulsion systems, are well placed to take advantage of this kind of adaptive testing.
Practical Recommendations for Applying the Model
- Calibrate the temperature reconstruction carefully. A regression on garbage is still garbage, and the largest source of error in the model is usually the translation from measured thermocouple readings to estimated contact temperature.
- Use a design of experiments that covers the full operating envelope, not only the nominal point. The model will be extrapolated in service, and it needs to be trained across the range it will be asked to score.
- Keep one physics-based simulation in the loop, even if it is run rarely. The simulation acts as a sanity check on the regression when the test data are sparse or when a new operating regime is introduced.
- Track the model performance in service, not only in the test cell. Drift in sensor calibration, changes in lubricant batch, and surface wear can all move the model away from its training distribution, and periodic re-fitting is part of the deal.
- Document the feature engineering transparently. Future engineers, including those in Australian regulatory and certification roles, will want to be able to audit exactly how the model arrived at its score.
If you want to follow the work as it evolves, the project page is the central point of contact, and registered collaborators can pull raw test data, model coefficients, and validation reports from the members area. The team is open to dialogue with Australian engine programs and with the research groups that support them, and any engineer with a gear rig, a temperature trace, and a stubborn scuffing mystery is welcome to get in touch.