Modelling Gear Tooth Roughness And EHL Power Loss
A regression model linking gear tooth roughness to power loss in the elastohydrodynamic regime can give gearbox designers a practical way to connect surface metrology with propulsion performance. Instead of treating roughness as a pass-or-fail manufacturing result, the model estimates how texture, lubricant behaviour, load, speed and temperature combine to alter frictional losses at the tooth contact.
This approach fits the OPTIMIZE Project’s broader engineering method. Geared aircraft engines operate with demanding power density, high pitch-line velocity and narrow reliability margins, so a small change in mesh efficiency can affect thermal balance, fuel consumption, durability and certification evidence. For Australian manufacturers and research teams, a robust relationship between surface finish and loss can also help reduce expensive trial-and-error testing across a dispersed local supply chain.
Why Roughness Matters In Elastohydrodynamic Contacts
In an elastohydrodynamic lubrication, or EHL, contact, the mating gear teeth are separated by a very thin lubricant film. The pressure is high enough to deform the tooth surfaces elastically, while lubricant viscosity rises sharply within the contact. This creates a load-carrying film, yet the teeth still experience rolling and sliding, producing frictional power loss.
Surface roughness modifies that film in several ways. A roughness peak can locally reduce film thickness, increase asperity interaction and disturb the pressure distribution. Wavelength also matters. A fine texture may be smoothed by the lubricant film, whereas longer waviness can influence the whole contact patch. The relevant variable is therefore not simply Ra, the arithmetic average roughness, but a group of parameters describing height, spacing, directionality and spectral content.
A regression model should therefore represent roughness as a measurable cause of loss rather than as a single quality-control label. Candidate responses include mesh friction torque, temperature rise, coefficient of friction and total gearbox power loss. The response may be normalised by transmitted torque or input power so that tests conducted at different operating points can be compared fairly.
A sound model can also prevent a common interpretation error. If a gearbox loses more power after a manufacturing change, roughness may be involved, but the cause could also be altered tooth lead, profile deviation, lubricant aeration, bearing drag or thermal expansion. The regression must separate these effects through controlled design of experiments rather than assign every efficiency change to surface texture.
Selecting Surface And Operating Variables
The strongest predictors will usually combine manufacturing measurements with operating conditions. Tooth flank roughness can be measured before assembly using a stylus instrument, optical profiler or areal interferometer. Measurements should record the direction of lay because roughness aligned with sliding can behave differently from roughness crossing the direction of motion.
The model should also include composite roughness for the contacting pair. A practical form is the root-sum-square value of the pinion and gear surface heights, although separate terms may be retained when the two components use different finishing processes. Skewness, kurtosis, peak density, autocorrelation length and filtered roughness can reveal whether isolated peaks or repeating patterns are driving the result.
The relationship between roughness and friction is strongly dependent on film thickness. Speed, load, oil temperature, viscosity grade and lubricant supply should therefore be recorded at the same time as torque loss. A useful dimensionless indicator is the lambda ratio, which compares central film thickness with composite surface roughness. Low lambda values indicate greater asperity interaction and a greater chance that roughness will influence power loss.
Bearing behaviour should be monitored as a possible confounding factor. Raceway waviness at particular orders can transmit vibration into the gear mesh and alter measured torque or temperature. The OPTIMIZE Project’s discussion of raceway waviness effects is relevant because surface geometry elsewhere in the gearbox can appear in the response used to fit the roughness model.
A useful experimental dataset may include:
- Roughness amplitude, skewness, kurtosis and autocorrelation length for each tooth flank
- Pinion and gear rotational speed, transmitted torque and sliding-to-rolling ratio
- Oil inlet temperature, bulk temperature, viscosity and lubricant flow rate
- Tooth profile and lead deviations, alignment error and mesh stiffness indicators
- Measured input and output torque with uncertainty estimates for each instrument
- Gearbox housing temperature, vibration order content and bearing drag
Turning Test Results Into A Predictive Equation
A first model can use a response such as frictional power loss divided by transmitted power. A general expression might include linear terms for composite roughness and lambda ratio, plus interaction terms for roughness with sliding speed and temperature. Polynomial terms can capture curvature where loss rises sharply after the lubricant film becomes too thin.
One possible structure is:
[ P_\mathrm{loss}^\ast = \beta_0+\beta_1R_q+\beta_2\lambda^{-1}+\beta_3V_s+ \beta_4T_o+\beta_5R_qV_s+\beta_6R_q\lambda^{-1}+\varepsilon ]
Here, (P_\mathrm{loss}^\ast) is normalised power loss, (R_q) is composite root-mean-square roughness, (\lambda) is the film parameter, (V_s) is sliding velocity, (T_o) is oil temperature and (\varepsilon) is the residual error. In practice, transformed variables such as logarithms may be more suitable if the loss response is skewed or changes over several orders of magnitude.
The equation should remain interpretable. A highly flexible machine-learning model may fit the laboratory data closely while failing when a new gear batch, lubricant or load cycle is introduced. Ridge regression, partial least squares, Gaussian-process regression or a carefully constrained gradient-boosting model can be considered, but each should be compared with a simpler response-surface model.
| Model element | Engineering meaning | Recommended treatment | Main risk |
|---|---|---|---|
| Composite roughness | Combined height variation of mating flanks | Use measured areal or profile parameters with uncertainty | Reducing texture to Ra alone |
| Lambda ratio | Film thickness relative to surface texture | Include as a primary operating variable | Calculating film thickness from incorrect viscosity data |
| Sliding velocity | Frictional work generated during tooth sliding | Add linear and interaction terms | Ignoring changes across the mesh |
| Oil temperature | Controls viscosity and film formation | Measure at inlet and near the mesh | Confusing thermal equilibrium with inlet conditions |
| Torque and speed | Define load and transmitted power | Cover the complete intended operating envelope | Fitting only one design point |
| Manufacturing variation | Captures batch-to-batch and tooth-to-tooth effects | Use mixed-effects or hierarchical terms | Treating repeated tests as independent |
| Residual uncertainty | Represents instrument and test variability | Report prediction intervals, not only R² | Presenting a precise-looking but fragile estimate |
The coefficients should be estimated using a designed experiment rather than a random collection of tests. A factorial or response-surface design can vary torque, speed, temperature, lubricant flow and surface finish while keeping the number of runs manageable. Replicates at the centre point help identify repeatability, while runs near the operating envelope reveal whether the roughness effect becomes nonlinear.
Australian test programmes often need to make each run count. A component may be manufactured in Melbourne or Geelong, finished by a specialist supplier interstate and tested at a university or research facility several hours away. That arrangement makes disciplined test planning especially valuable: replacing a damaged aerospace gear set is costly, and transport delays can interrupt a narrow test window.
Checking The Model Against Physical Behaviour
Statistical fit is necessary but insufficient. A model with a high coefficient of determination can still violate tribological behaviour by predicting lower loss as roughness increases under a low-film condition. Engineers should inspect coefficient signs, partial dependence plots and residuals against speed, load, temperature and surface condition.
Validation should be performed by grouping data according to physical sources of variation. For example, all tests from one manufacturing batch should not be split randomly between training and validation sets, because that can make the model appear more accurate than it is. Leave-one-batch-out validation, leave-one-lubricant-out validation and tests on a new gear pair provide a stronger indication of industrial usefulness.
The model should also be checked against energy balance. Input torque and speed define mechanical power entering the gearbox, while output torque, speed, bearing losses, seal losses and gear mesh losses account for the measured result. If the roughness term predicts a change larger than the total unexplained loss, the model is capturing an instrument problem or an omitted variable rather than a physical effect.
Useful validation evidence includes:
- Repeatability tests on the same gear pair at stable temperature
- Independent tests using a new surface-finish batch
- Inter-laboratory checks for roughness measurement and torque calibration
- Load and speed sweeps covering low, nominal and high operating points
- Microscopic inspection after testing to identify scuffing or polishing
- Comparison between predicted loss, thermal rise and measured efficiency
Surface characterisation deserves particular care. Two tooth flanks can have the same Ra while producing different friction because one contains directional grinding marks and the other has isolated sharp peaks. A stylus trace may also miss areal features or fail to capture the portion of the flank that carries the highest load. Measurement location, filtering cut-off and instrument resolution should be stored with every observation in the regression dataset.
For Australian certification and production, traceability is commercially important. Local suppliers may operate in relatively small batches rather than the high-volume conditions found in major overseas aerospace hubs. A model that records measurement uncertainty, tool condition and process route can help a supplier demonstrate consistent capability without promising a level of precision the manufacturing process cannot sustain.
Using The Results In Gearbox Design
Once validated, the regression model can become a design and manufacturing decision tool. It can estimate the efficiency penalty associated with a proposed surface-finish range, identify where polishing effort has the greatest value and support tolerance allocation between roughness, profile error and alignment. It can also be embedded in a broader gearbox simulation that includes thermal, dynamic and structural calculations.
The output should be expressed in engineering terms. A designer may need the predicted watts lost at a cruise condition, the expected temperature increase during take-off, or the probability that a roughness limit will push the gearbox beyond its thermal budget. Confidence and prediction intervals should accompany every estimate, particularly when the model is extrapolated beyond the test envelope.
A cost-benefit comparison can then distinguish between useful finishing and unnecessary refinement. If reducing roughness from one level to a finer level saves only a small amount of power but increases manufacturing time substantially, the change may not be justified. If a moderate surface improvement prevents local heating under high sliding conditions, the same process change may be worthwhile even when average efficiency gains look modest.
Communication matters as much as the mathematics. An engineer explaining the model to a production manager may use watts, efficiency percentage and inspection time, while a tribologist will focus on lambda ratio and asperity interaction. Public-facing technical material can also benefit from plain-language explanations; even a general digital product such as an expanding casino interface illustrates how quickly users lose confidence when an important system response is difficult to interpret. The gearbox model should make its inputs, limits and uncertainty equally clear.
For operators and maintainers, the model may support condition monitoring. A gradual rise in measured loss, housing temperature or vibration could indicate surface distress, lubricant degradation or a change in tooth contact. Roughness itself is not always measurable in service, so indirect indicators must be used carefully and linked back to physical inspection during overhaul.
Australian operating conditions can add practical constraints. Aircraft and test assets may move between humid coastal facilities around Sydney or Brisbane and hot, dry inland environments, with different effects on cooling, storage and lubricant handling. Remote maintenance locations need procedures that are robust when specialist metrology is not immediately available. A model based on a well-defined measurement chain can help decide which checks must occur locally and which can wait for a central laboratory.
Moving From Regression To Qualification
The next step is to connect the roughness-loss relationship with design-of-experiments studies covering gearbox architecture, lubrication, manufacturing tolerance and operating duty. The result should be a calibrated engineering surrogate, not a replacement for physical testing. Simulation can screen candidate designs, while rig tests confirm the predicted trends under realistic speed, load and thermal conditions.
Qualification plans should preserve the distinction between efficiency testing and durability testing. A smooth tooth surface may reduce friction while a different finishing process changes residual stress, micropitting resistance or tooth-root performance. The preferred manufacturing window must satisfy the whole gearbox requirement, including noise, wear, fatigue life, weight and power density.
Commercial decisions also benefit from transparent uncertainty. A procurement team assessing a new Australian supplier may compare predicted power loss, process capability and inspection cost rather than selecting solely on nominal roughness. If a project includes external financial or operational comparisons, even a phrase such as cash-out conditions demonstrates why definitions matter: an apparently simple result depends on the conditions attached to it. For gearbox work, those conditions include speed, torque, oil temperature, measurement filter and test duration.
The OPTIMIZE Project’s engineering focus provides a suitable framework for this integration. Design-of-experiments methods identify influential factors, simulation explains the contact mechanics, tolerance analysis represents manufacturing variation, and physical testing checks whether the predicted efficiency and durability improvements survive in a real gearbox.
A clear implementation sequence is:
- Define the loss response, measurement uncertainty and operating envelope
- Characterise tooth surfaces with repeatable profile and areal methods
- Build a designed test matrix covering roughness, load, speed and temperature
- Fit interpretable regression models before testing more complex algorithms
- Validate by manufacturing batch, lubricant and gear pair rather than random rows
- Feed the prediction intervals into design, inspection and qualification decisions
Researchers, suppliers and operators can contribute useful evidence by sharing carefully documented test conditions and failure observations. The project’s contact information provides a route for engaging with the wider research and engineering work. With consistent data, a roughness-based regression model can turn a difficult tribological variable into a practical control point for quieter, lighter and more efficient geared aircraft engines.
The most valuable outcome is a defensible link between what a manufacturing instrument measures and what an aircraft gearbox consumes in operation. When that link is tested across realistic Australian conditions, manufacturing variation and propulsion duty cycles, surface finish becomes part of a quantified design strategy rather than an isolated workshop specification.