Killing the Thermal Solver at Peak Torque: Fault Injection for Coupled Gearbox Co-Simulation

The failures that matter in a coupled gearbox model do not appear at nominal torque. They show up when mesh friction heat, tooth deflection and oil film thickness start feeding one another, and that is the regime where one partition of the co-simulation is most likely to stall. A qualification plan that only proves the coupled run converges under load has proved the easy half. The hard half is showing what the orchestrator does when the thermal partition stops answering mid-ramp, and whether what survives is still physics or just numbers.

Why load hides the fault

Mesh friction couples the structural and thermal partitions through a coefficient that depends on sliding velocity, contact pressure and bulk temperature, so heat generated at the flank changes clearance, and clearance changes how load is shared between teeth. At light load this loop is weak, and a partition can lag a few steps without visible consequence. Under a torque ramp the loop stiffens, the thermal solver's implicit iterations lengthen, and a missed deadline becomes likely exactly where the physics is most sensitive. If the orchestrator responds by holding the last thermal value, the friction coefficient freezes at a cooler state, predicted flash temperature drops, and the model quietly underpredicts scuffing risk. Nothing crashes. The dashboard stays green, and that is the failure.

The test: remove a partition where it hurts

The test that catches this drives the full coupled model through a torque ramp into the scuffing-relevant load range, and at the instant friction heat flux peaks the harness terminates the thermal unit outright, not by throttling it but by removing it from the master algorithm. The communication step stays fixed at 0.5 ms for the whole run, so the takeover is judged at the same temporal resolution as the reference. The orchestrator has one macro step to detect the loss, switch in a reduced-order thermal surrogate trained on the same operating envelope, and keep the structural partition advancing without a jump in contact force. The verdict comes from a twin run in which every partition stays alive, compared on flank temperature trajectory, transmission error and, above all, the energy balance across the coupling interface. A run that survives but drifts from its twin has not passed. It has only avoided crashing.

The decision rule

If the energy residual accumulated at the thermal-structural interface after takeover exceeds the friction heat that interface should have carried over the same window, the run is declared invalid rather than degraded, and the surrogate is not allowed to sign off a load case it cannot conserve.

What the plan still has to settle

Co-simulation interface standards such as FMI and SSP give the harness a clean way to terminate and replace a unit, but they say nothing about whether the replacement is physically admissible, so that judgement stays with the qualification plan. The surrogate's envelope is the weak point: train it on ramps that never reached the scuffing regime and it will extrapolate confidently and wrongly. Covering that envelope properly costs a real share of the original high-fidelity campaign, roughly a fifth as an estimate, and that budget competes with every other load case. So the next decision is not whether to run the fault-injection test but which partition gets a qualified surrogate first, and the honest answer is whichever one stalls closest to the physics you are trying to certify.

Why not just retry the failed partition instead of switching to a surrogate?

At peak load a retry usually rebuilds the same stiff system that stalled in the first place, and every retried step stretches wall-clock time enough to desynchronise the real-time partitions. The surrogate keeps the timeline intact and makes the lost fidelity measurable through the energy residual, which a silent retry loop never does.

Back to MechSim · Domains

← Back to MechSim
Engineering & Architecture Brief

Killing the Thermal Solver at Peak Torque: Fault Injection for Coupled Gearbox Co-Simulation

The failures that matter in a coupled gearbox model do not appear at nominal torque. They show up when mesh friction heat, tooth deflection and oil film thickness start feeding one another, and that is the regime where one partition of the co-simulation is most likely to stall. A qualification plan that only proves the coupled run converges under load has proved the easy half. The hard half is showing what the orchestrator does when the thermal partition stops answering mid-ramp, and whether what survives is still physics or just numbers.

Why load hides the fault

Mesh friction couples the structural and thermal partitions through a coefficient that depends on sliding velocity, contact pressure and bulk temperature, so heat generated at the flank changes clearance, and clearance changes how load is shared between teeth. At light load this loop is weak, and a partition can lag a few steps without visible consequence. Under a torque ramp the loop stiffens, the thermal solver's implicit iterations lengthen, and a missed deadline becomes likely exactly where the physics is most sensitive. If the orchestrator responds by holding the last thermal value, the friction coefficient freezes at a cooler state, predicted flash temperature drops, and the model quietly underpredicts scuffing risk. Nothing crashes. The dashboard stays green, and that is the failure.

The test: remove a partition where it hurts

The test that catches this drives the full coupled model through a torque ramp into the scuffing-relevant load range, and at the instant friction heat flux peaks the harness terminates the thermal unit outright, not by throttling it but by removing it from the master algorithm. The communication step stays fixed at 0.5 ms for the whole run, so the takeover is judged at the same temporal resolution as the reference. The orchestrator has one macro step to detect the loss, switch in a reduced-order thermal surrogate trained on the same operating envelope, and keep the structural partition advancing without a jump in contact force. The verdict comes from a twin run in which every partition stays alive, compared on flank temperature trajectory, transmission error and, above all, the energy balance across the coupling interface. A run that survives but drifts from its twin has not passed. It has only avoided crashing.

The decision rule

If the energy residual accumulated at the thermal-structural interface after takeover exceeds the friction heat that interface should have carried over the same window, the run is declared invalid rather than degraded, and the surrogate is not allowed to sign off a load case it cannot conserve.

What the plan still has to settle

Co-simulation interface standards such as FMI and SSP give the harness a clean way to terminate and replace a unit, but they say nothing about whether the replacement is physically admissible, so that judgement stays with the qualification plan. The surrogate's envelope is the weak point: train it on ramps that never reached the scuffing regime and it will extrapolate confidently and wrongly. Covering that envelope properly costs a real share of the original high-fidelity campaign, roughly a fifth as an estimate, and that budget competes with every other load case. So the next decision is not whether to run the fault-injection test but which partition gets a qualified surrogate first, and the honest answer is whichever one stalls closest to the physics you are trying to certify.

Technical FAQ: Why not just retry the failed partition instead of switching to a surrogate?

At peak load a retry usually rebuilds the same stiff system that stalled in the first place, and every retried step stretches wall-clock time enough to desynchronise the real-time partitions. The surrogate keeps the timeline intact and makes the lost fidelity measurable through the energy residual, which a silent retry loop never does.

Related Architecture Specifications (DIGITALTWINS): neuraldigitaltwins.com · toursimulation.com · isim2real.com

More Technical Articles on mechsim.com

Synthetic Bearing Spall Growth and Acoustic Emission Vibration Benchmarks →Thermal Choking and Churning Power Loss in High-Speed Dual-Clutch Transmissions →Reference Architecture for Gearbox Thermal-Structural Co-Simulation: Multirate Coupling with FMI 3.0, SSP 1.0 and ISO 6336 →