Engineers rarely get more than one trustworthy data point. That point is often a single rig test of a friction clutch, or a single expensive fluid simulation of how a rotating drum is cooled. They then have to decide whether a slightly different design will run hotter or cooler, and whether it is closer to or further from surface thermal cracking. This proposal asks whether one case, well understood, can predict the direction of change for a whole neighborhood of designs. The goal is not to predict exact temperatures. The method should also say explicitly where its predictions stop being trustworthy and which single extra simulation would resolve the doubt.
The tool is qualitative causal modeling: a branch of AI that represents a physical system as signs and monotonic relations ("heat input rises with sliding speed") instead of exact equations. Its known weakness is the competing-influence problem. When frictional heating rises and convective cooling also rises, the qualitative algebra cannot tell which wins. Frictional thermal cracking in a ventilated rotating drum is a clean example of that problem. Our central idea is to resolve each unresolved competition with dominance evidence taken from the one high-fidelity case. That case is a conjugate heat-transfer simulation, in which fluid and solid heat flow are solved together. The fluid side uses the RNG $k$-$\epsilon$ turbulence model, a two-equation model with a strain-rate correction suited to swirling, rotating flows. The qualitative model then carries this evidence across finite perturbations and regime boundaries. It flags the perturbations where the evidence can no longer be transported, and those flags become the schedule for the next simulations.
Core Claim
A hybrid of qualitative comparative analysis and simulation-derived dominance labels gives correct signs of change in peak surface temperature and cracking margin over a family of perturbed friction-drum designs. It does this from one or a few high-fidelity cases, under perturbations large enough to break local linearization. Its "ambiguous" flags should concentrate on the perturbations where the true sign actually flips. The specific novelty is twofold. First, the qualitative model is used to turn an $N=1$ engineering case into a set of bounded directional claims. Second, the qualitative model's own indeterminacy is used to choose which additional case to run.
Where This Sits
Qualitative reasoning supplies the method. Weld's Comparative Analysis infers how a behavior shifts under a parameter perturbation. It reports correct answers on a spring-block oscillator but is explicitly incomplete when opposing perturbations act together. We fill that gap with numerical dominance evidence instead of the asymptotic "exaggeration" device it uses. The Limits of Qualitative Simulation gives two things we rely on: the QSIM constraint vocabulary (monotonic relations, landmarks) and the diagnosis that net-sign ambiguity causes spurious branching. We treat that branching as a signal rather than a defect. Farquhar's A Qualitative Physics Compiler compiles reusable model fragments into qualitative equations and was benchmarked on conductive heat transfer between three bodies. We reuse the model-fragment idea, but QPC has no friction, hysteresis, turbulent convection, or cracking. Bylander's The Qualitative Difference Resolution Rule and the QDOCS diagnoser in Qualitative Multiple-Fault Diagnosis of Continuous Dynamic Systems Using Behavioral Modes show that qualitative models scale to real lumped fluid networks. Neither touches continuum turbulence or contact heating, and neither uses a qualitative model to generalize from one case. None of these qualitative reasoning papers couples to a CFD case, so the bridge we propose is absent from the file.
Several papers motivate the problem. Experimental Analysis on High-Rated High-Elastic Clutch Coupling ties drum thermal cracking to the ratio of absorbed friction energy to sliding time. It also found that wear in the 10,000 ps full-scale trial departed from the 500 ps scale model, because of initial wear and unsteady engagement that its thermal calculation could not capture. That is a textbook failure of extrapolating from one case. Velocity-Dependent Dynamics of Friction and Wear maps friction across velocity regimes in silicon carbide: strengthening, weakening at a critical velocity near 32 m/s, viscous strengthening, and flash heating above 800 m/s. These regimes are qualitative landmarks where the sign of $\partial \mu / \partial v$ (how the friction coefficient $\mu$ changes with sliding speed $v$) changes. A linearization at one point will miss them. Axial Air-cooled Road Wheel Design and Advanced Calculation Model of V-Belt Temperature Field add internal hysteretic heating to the competition. The V-belt study reports over 65% of total heat coming from hysteresis.
Other papers constrain the design. On the turbulence side, Numerical Simulation of Flow Field of DN2400 Hedge Dissipator is an RNG $k$-$\epsilon$ study of a single geometry. It needed 3.574 million cells and offers no way to carry its findings to other sizes, which is the cost that makes re-running cases impractical. Research on the parameters of turbulence model and modeling of equilibrium atmosphere boundary layer in CWE shows that a closure coefficient ($C_\mu$ from 0.09 to 0.04) can change conclusions. The choice of turbulence closure is therefore itself an $N=1$ choice, and we treat it as one more perturbation axis.
Research Program
We build a qualitative model of a ventilated rotating friction drum from model fragments. The fragments cover:
- interfacial friction heating, with velocity-regime landmarks taken from the SiC regime map;
- hysteretic bulk heating in an elastomeric element;
- conduction into the drum;
- forced convection to the cooling air;
- thermal gradient as the cracking driver, with a landmark at the cracking threshold.
The central competition is written as one sign relation:
$$\operatorname{sign}\!\left(\frac{dT_s}{dt}\right) = \operatorname{sign}\!\left(q_f + q_h - h A (T_s - T_\infty)\right)$$
Here $T_s$ is peak surface temperature, $q_f$ and $q_h$ are frictional and hysteretic heat input, $h$ is the convective coefficient, $A$ the cooled area, and $T_\infty$ the coolant temperature. A perturbation such as raising rotation speed increases both $q_f$ and $h$. Comparative analysis alone returns "unknown" for this case.
The high-fidelity case is an RNG $k$-$\epsilon$ conjugate heat-transfer simulation of one baseline design. From it we extract local dominance labels: the sign of each competing term's sensitivity to each design parameter. We obtain these by finite differences and also record their magnitudes as a margin. Each label is attached to the qualitative region, meaning the set of landmark intervals, in which it was measured. The comparative-analysis engine then propagates signs to a perturbed design under one rule. A label is reused only if the perturbed design stays in the same qualitative region and the recorded margin exceeds a threshold. Otherwise the prediction is marked ambiguous.
When ambiguous predictions accumulate, the system picks the next case to simulate. It chooses the perturbed design that would resolve the most ambiguous branches at once. This is a discrete set-cover selection over branches, not a learned acquisition function. The deliverables are three: the fragment library, the propagation engine, and a protocol that reports, for any proposed design change, a sign, a regime it is valid in, and a confidence derived from the margin.
Baselines
The naive reference point is direct transfer of the single case's empirical rule. Following the clutch coupling paper, it predicts that cracking risk moves with the friction-energy-to-sliding-time ratio and nothing else. All cooling and regime effects are treated as unchanged. It uses the same baseline case and costs nothing extra. Beating it means higher sign accuracy on perturbations that alter cooling geometry or cross a velocity landmark. Matching it elsewhere is acceptable.
The strongest competitor for directional questions is local sensitivity analysis at the baseline case. It takes the same finite-difference derivatives we compute and extrapolates their signs linearly to every perturbed design. It is strong because it uses exactly our information minus the qualitative structure. It is therefore also the ablation that isolates our key ingredient, which is regime-aware transport and abstention. It gets the identical simulation budget. We count it as beaten only if the hybrid is more accurate on perturbations beyond some fraction of the parameter range, or across a landmark. We also require that the hybrid's abstentions recover a measurable share of the sign flips that sensitivity analysis gets wrong.
The strongest data-driven competitor is a Gaussian-process surrogate with active learning, a statistical curve-fitter that also reports its own uncertainty. It is given the same number of high-fidelity runs and the same input parameters, with hyperparameters tuned by marginal likelihood. Its predictive sign and uncertainty are compared with ours. We also report dimensional scaling through Biot and Péclet-type groups as the classical engineering baseline.
The qualitative-only ablation runs comparative analysis without dominance labels. It shows how much the labels contribute.
For reference, no reported result found on Lacuna for sign-of-change prediction of frictional thermal cracking under design perturbation. The nearest qualitative-reasoning figure is the 77% correct-diagnosis rate of QDOCS on the Space Shuttle reaction control system, reported in Qualitative Multiple-Fault Diagnosis of Continuous Dynamic Systems Using Behavioral Modes. That is a different task and serves as context only, not a target.
Why not just…?
Why not just run local sensitivity analysis at the base case? This is the real threat. If every perturbation of interest stays inside one regime, linear signs will be as accurate as ours and simpler. Our claim rests on two assumptions. One is that thermal cracking problems routinely cross landmarks, such as the velocity-weakening transition, the onset of flash heating, or the cracking threshold itself. The other is that the competing-term margins are often small, so finite perturbations flip the sign. The decisive experiment stratifies test perturbations by whether they cross a landmark and by perturbation size. If the hybrid shows no advantage in the crossing stratum, we will report that the qualitative layer adds only abstention value, and narrow the claim to that.
Why not just fit a surrogate model with active learning? A Gaussian process or a neural operator like the one in PDE-Transformer: Efficient and Versatile Transformers for Physics Simulations can interpolate well. That paper reports a 42% average nRMSE improvement over FNO, but on synthetic PDEs with plentiful training data. At two to five high-fidelity runs, a surrogate's uncertainty reflects distance from data, not physical regime changes. The matched-budget comparison will show whether this matters. If the surrogate matches our sign accuracy at three runs, the qualitative layer's value reduces to interpretability, and we will say so.
Why not just use dimensionless scaling laws? Dimensional analysis is itself a qualitative method and handles geometric similarity well. The clutch paper's 500 ps versus 10,000 ps discrepancy shows where it fails. Initial wear and unsteady engagement change which mechanism dominates, and similarity does not capture that. We include scaling as a baseline and expect it to do well on pure size perturbations and poorly on cooling-geometry and regime-crossing changes.
Proposed Experiments
The ground-truth set is a perturbation grid around one baseline ventilated drum. It varies rotation speed, cooling-channel area, lining conductivity, engagement time, and hysteretic loss fraction. Each grid point is run as an RNG $k$-$\epsilon$ conjugate heat-transfer simulation. We score four things:
- sign accuracy of change in peak surface temperature and in cracking margin over all baseline-to-perturbed pairs;
- coverage, meaning the fraction of pairs on which the method commits to a sign;
- abstention precision, meaning the fraction of ambiguous flags that correspond to true sign reversals or regime crossings;
- the number of extra high-fidelity runs needed to reach a fixed sign accuracy, comparing set-cover selection against random and surrogate-driven selection.
Three further checks probe robustness. To test whether conclusions survive a change of turbulence closure, we repeat a subset under SST $k$-$\omega$ and standard $k$-$\epsilon$ with recalibrated $C_\mu$. A predicted sign counts as robust only if all closures agree. For out-of-family transfer, we apply the library unchanged to an axial air-cooled road-wheel geometry patterned on the road-wheel paper, using its own single base case. As a historical sanity check, we ask whether the model, given the 500 ps clutch case, flags the 10,000 ps extrapolation as ambiguous rather than confidently wrong. This check is one data point and will be reported as anecdote, not evidence.
The proposal succeeds if the hybrid beats local sensitivity analysis in the landmark-crossing stratum and its abstentions catch most sign flips. It also succeeds, in a narrower form, if it reaches the surrogate's accuracy with fewer runs. If neither holds, the result is still useful: a clear negative finding that, for this class of problems, a single well-chosen derivative is all one case can teach.
References
- Comparative Analysis
- The Limits of Qualitative Simulation
- A Qualitative Physics Compiler
- The Qualitative Difference Resolution Rule
- Qualitative Multiple-Fault Diagnosis of Continuous Dynamic Systems Using Behavioral Modes
- Experimental Analysis on High-Rated High-Elastic Clutch Coupling
- Velocity-Dependent Dynamics of Friction and Wear
- Axial Air-cooled Road Wheel Design
- Advanced Calculation Model of V-Belt Temperature Field
- Numerical Simulation of Flow Field of DN2400 Hedge Dissipator
- Research on the parameters of turbulence model and modeling of equilibrium atmosphere boundary layer in CWE
- PDE-Transformer: Efficient and Versatile Transformers for Physics Simulations