Killing the Thermal Solver at Full Torque: Qualifying Graceful Degradation in Coupled Gearbox Co-Simulation

A coupled gearbox co-simulation can pass every nominal regression and still fail the first time the thermal solver stalls at rated torque, because the coupling only carries real information when the mesh is hot. At light load the friction heat is small and oil viscosity barely moves, so a frozen temperature field looks exactly like a live one. Under load the same frozen field tells the mesh friction model that the oil is thicker than it really is. The structural side also never sees the thermal growth that is closing the backlash.

Why nominal runs never see it

The thermal and structural solvers exchange state at a fixed coupling interval, typically a 2 ms macro step for a mesh dynamics model running against a lumped or reduced thermal network. Between exchanges, each side runs on what it last received. If the thermal unit hangs, returns late, or quietly repeats its previous output, the master algorithm usually extrapolates, and at low torque that extrapolation is harmless because nothing is changing. At rated torque, friction heat at the flank rises with sliding velocity and load. Viscosity drops, the film thins, the friction coefficient climbs, and that feedback loop is exactly the physics the stale field cuts off. The failure is a co-simulation that stays numerically stable, produces smooth plots, and underpredicts flash temperature right in the regime the qualification exists to cover.

The test: fail the thermal unit at peak load

Run the duty cycle until the gearbox reaches thermal soak at rated torque. Only then inject the fault, and inject it three ways in separate runs: a hard stall where the thermal unit never returns, a silent repeat where it keeps sending its last valid state, and a late return where valid data arrives several macro steps behind. Each run is compared against two references, an unfaulted coupled run and a run where thermal coupling is disabled from the start, so you can tell whether the faulted run falls back toward a known bound or drifts somewhere in between. To pass, the master has to detect the fault within one macro step and switch to a declared degraded mode that assumes worst-case viscosity at the maximum allowed sump temperature. It also has to tag every downstream output as degraded. As an estimate, a stalled thermal unit at rated torque shows up in contact temperature within tens of seconds of simulated time, much shorter than any nominal soak, so the silent repeat case is the one most likely to slip past a reviewer, and it gets the strictest check.

The decision rule for the master is to reject the step and never extrapolate whenever the interface energy residual over a macro step exceeds the mesh friction heat the thermal side should have absorbed in that same step, because at that point the coupling is creating or destroying more energy than the physics you are trying to resolve.

What the safety case asks of the fallback

Under functional safety standards such as ISO 26262 and IEC 61508, simulation results used as verification evidence inherit the tool confidence argument, so the degraded mode has to be qualified behaviour, not an accident of the master algorithm. The fault injection runs above become part of that argument, since they show the fallback is conservative, detectable and bounded. The cost is that a conservative fallback over-predicts scuffing risk, and a design that is fine in reality can fail a degraded run. The test lead then has to decide whether a degraded result can count as evidence at all, or whether any run that enters degraded mode is thrown out and repeated with the thermal unit healthy, which costs compute but keeps the evidence chain clean.

Why not let the master extrapolate temperature for a few steps instead of rejecting them?

Extrapolation assumes the thermal trend you last saw will continue, but at rated torque the trend is driven by a viscosity and friction feedback loop that extrapolation cannot represent. A rejected step costs you throughput. An extrapolated step that hides a thermal runaway costs you the validity of the whole run, and nothing in the output warns you.

Back to MechSim · Domains

← Back to MechSim
Engineering & Architecture Brief

Killing the Thermal Solver at Full Torque: Qualifying Graceful Degradation in Coupled Gearbox Co-Simulation

A coupled gearbox co-simulation can pass every nominal regression and still fail the first time the thermal solver stalls at rated torque, because the coupling only carries real information when the mesh is hot. At light load the friction heat is small and oil viscosity barely moves, so a frozen temperature field looks exactly like a live one. Under load the same frozen field tells the mesh friction model that the oil is thicker than it really is. The structural side also never sees the thermal growth that is closing the backlash.

Why nominal runs never see it

The thermal and structural solvers exchange state at a fixed coupling interval, typically a 2 ms macro step for a mesh dynamics model running against a lumped or reduced thermal network. Between exchanges, each side runs on what it last received. If the thermal unit hangs, returns late, or quietly repeats its previous output, the master algorithm usually extrapolates, and at low torque that extrapolation is harmless because nothing is changing. At rated torque, friction heat at the flank rises with sliding velocity and load. Viscosity drops, the film thins, the friction coefficient climbs, and that feedback loop is exactly the physics the stale field cuts off. The failure is a co-simulation that stays numerically stable, produces smooth plots, and underpredicts flash temperature right in the regime the qualification exists to cover.

The test: fail the thermal unit at peak load

Run the duty cycle until the gearbox reaches thermal soak at rated torque. Only then inject the fault, and inject it three ways in separate runs: a hard stall where the thermal unit never returns, a silent repeat where it keeps sending its last valid state, and a late return where valid data arrives several macro steps behind. Each run is compared against two references, an unfaulted coupled run and a run where thermal coupling is disabled from the start, so you can tell whether the faulted run falls back toward a known bound or drifts somewhere in between. To pass, the master has to detect the fault within one macro step and switch to a declared degraded mode that assumes worst-case viscosity at the maximum allowed sump temperature. It also has to tag every downstream output as degraded. As an estimate, a stalled thermal unit at rated torque shows up in contact temperature within tens of seconds of simulated time, much shorter than any nominal soak, so the silent repeat case is the one most likely to slip past a reviewer, and it gets the strictest check.

The decision rule for the master is to reject the step and never extrapolate whenever the interface energy residual over a macro step exceeds the mesh friction heat the thermal side should have absorbed in that same step, because at that point the coupling is creating or destroying more energy than the physics you are trying to resolve.

What the safety case asks of the fallback

Under functional safety standards such as ISO 26262 and IEC 61508, simulation results used as verification evidence inherit the tool confidence argument, so the degraded mode has to be qualified behaviour, not an accident of the master algorithm. The fault injection runs above become part of that argument, since they show the fallback is conservative, detectable and bounded. The cost is that a conservative fallback over-predicts scuffing risk, and a design that is fine in reality can fail a degraded run. The test lead then has to decide whether a degraded result can count as evidence at all, or whether any run that enters degraded mode is thrown out and repeated with the thermal unit healthy, which costs compute but keeps the evidence chain clean.

Technical FAQ: Why not let the master extrapolate temperature for a few steps instead of rejecting them?

Extrapolation assumes the thermal trend you last saw will continue, but at rated torque the trend is driven by a viscosity and friction feedback loop that extrapolation cannot represent. A rejected step costs you throughput. An extrapolated step that hides a thermal runaway costs you the validity of the whole run, and nothing in the output warns you.

Related Architecture Specifications (DIGITALTWINS): supplysimulator.com · mechanicsim.com · regensim.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 →Killing the Thermal Solver at Peak Torque: Fault Injection for Coupled Gearbox Co-Simulation →Reference Architecture for Gearbox Thermal-Structural Co-Simulation: Multirate Coupling with FMI 3.0, SSP 1.0 and ISO 6336 →