Learning What Ultrasound Adds: Hypernetwork Corrections to a Passive Heat-Exchanger Surrogate, Trained and Tested on a Physical Rig

Heat exchangers can be made more efficient by combining two tricks. A fixed insert, such as a twisted strip inside the tube, swirls the flow. Vibrating actuators, such as ultrasonic transducers clamped to the tube, break up the thin, slow fluid layer at the wall. Experiments show the combination can raise heat transfer by close to 30% while lowering friction. The trouble is that nobody can predict the combined effect without building and testing each configuration. Ordinary fluid simulation handles the swirling insert reasonably well, but it cannot reliably resolve ultrasonic cavitation, the rapid growth and collapse of vapour bubbles that drives the active part of the enhancement. The effect is also non-monotonic: two transducers beat three, and 21 kHz beats 28 kHz.

We propose a hybrid model to fill this gap. A neural surrogate trained on simulations of the insert-only tube serves as a trustworthy "passive" model. A small hypernetwork (a network that outputs the weights, or weight corrections, of another network) takes the actuation settings as input: frequency, drive power, and where the transducers sit. It outputs a low-rank correction to the passive surrogate, and this correction is learned only from physical rig measurements. In two sentences: simulation supplies what simulation gets right (the swirl), and experiments teach a compact, actuation-conditioned correction for what simulation gets wrong (the ultrasound). We then test on held-out hardware configurations whether this correction predicts wall-temperature fields and efficiency better than the simpler ways engineers already calibrate models to data.

Core Claim

The claim is that the effect of ultrasonic forcing on a swirl-insert tube can be represented as a low-dimensional, smoothly actuation-dependent modification of the passive flow's temperature operator. If that is true, a hypernetwork can learn the modification from tens of rig runs rather than thousands of simulations. It will then predict spatially resolved wall temperatures, and from them the Nusselt number (dimensionless heat transfer), the friction factor, and the performance evaluation criterion (PEC), at actuation layouts it has never seen.

Concretely, let $G_\theta$ be the passive surrogate. It maps the Reynolds number $Re$, the applied heat flux $q$, and the inlet temperature to the unwrapped wall-temperature field $T(z,\phi)$ over axial position $z$ and azimuth $\phi$. Let $\mu$ be the actuation context. Each modulated layer $\ell$ then uses

$$W_\ell(\mu) = W_\ell^{\text{sim}} + U_\ell \,\mathrm{diag}\big(s_\ell(\mu)\big)\, V_\ell^{\top}, \qquad s_\ell(\mathbf{0}) = \mathbf{0},$$

where $W_\ell^{\text{sim}}$ are frozen weights fit to simulation, $U_\ell, V_\ell$ are shared learned bases of rank $r$, and $s_\ell(\mu)$ are singular values emitted by the hypernetwork. The constraint $s_\ell(\mathbf{0})=\mathbf{0}$ is enforced by gating the output with actuation power. It guarantees that the model with ultrasound switched off is exactly the passive surrogate.

What is new is not the use of a hypernetwork on physics. It is the division of labour: simulation-trained weights are held fixed, and a hypernetwork-generated correction captures the experimentally measured gap between simulation and reality, as a function of controllable actuator settings. It is also new to evaluate a dynamic-weight surrogate against physical thermo-fluid hardware.

Where This Sits in Prior Work

The application regime and the benchmark rig come from Study of ultrasonic influence on heat transfer and resistance performance of round tube with twisted belt. That study measured Nusselt number, friction and PEC for a 304 stainless tube with a 3D-printed twisted belt and piezoelectric transducers, over Re 5,000–18,000, at 21, 25 and 28 kHz, with one to three transducers. It is purely empirical. Its non-monotonic findings (two transducers best, 21 kHz best) motivate this proposal: they are exactly what a smooth scalar fit is likely to miss and what a context-conditioned model has to capture. We reuse its operating envelope and extend its instrumentation.

Three hypernetwork lines supply the mechanism, and none has touched experimental data. Extending Fourier Neural Operators for Modeling Parameterized and Coupled PDEs (hpFNO) has a hypernetwork emit layer-wise bias shifts for Fourier layers from physical parameters, and reports 55–72% lower error than coupled multiwavelet operators and HyperFNO on plasma, reaction–diffusion and flame simulations. Hypernetwork-based Meta-Learning for Low-Rank Physics-Informed Neural Networks (Hyper-LR-PINN) generates only the singular values of low-rank factored layers from PDE parameters, which is the form we adopt. It converges where MAML and Reptile fail, with about 30x fewer parameters. HyPINO generates whole physics-informed networks from coefficients and geometry, but is explicitly restricted to linear 2D problems. That restriction constrains us: we do not try to generate the full nonlinear thermo-fluid model from scratch, only a correction to one that already works. All three are trained and tested on synthetic solutions. Our correction instead has no simulation target at all, only measurements.

Scale-and-shift conditioning from other domains, such as layer-norm modulation in Deep Task-specific Bottom Representation Network for Multi-Task Recommendation and graph-conditioned scale/shift in Meta-Inductive Node Classification across Graphs, supplies the cheaper modulation form we use as an ablation.

Two surrogate papers constrain the choice of primary network. RecFNO reconstructs flow and heat fields from sparse sensors with Fourier operators, but reports that they blur abrupt local gradients near heat sinks. Ultrasonic enhancement is likely to be localized near transducers, so a Fourier primary may suffer the same way. A deep neural network surrogate modeling benchmark for temperature field prediction of heat source layout found U-Net with group normalization the most reliable field regressor, with under 3 ms inference per batch of 64. We therefore test both a Fourier and a U-Net primary. A Predictive Surrogate Model for Heat Transfer of an Impinging Jet on a Concave Surface shows that turbulence-model CFD validated against measured Nusselt distributions can train usable surrogates of convective heat transfer. It also shows that a mode-truncation (POD-LSTM) model loses accuracy over most of a pulsation cycle. That supports our reliance on CFD for the passive part and our preference for learned operators over mode truncation.

On validation methodology, Design and Experimental Verification of a PZT Pump with Streamlined Flow Tubes documents a characteristic simulation–prototype gap in a piezo-driven flow device: the flow-rate peak shifts from 10 to 11 Hz, and flow is 17.01 vs 18.75 mL/min at 120 V, attributed to bonding layers and connection losses. Mismatches of this kind, set by actuator mounting, are what our correction is meant to absorb. Quantifying heat transfer in DMD-based optoelectronic tweezers with infrared thermography supplies the infrared-thermography protocol for validating temperature fields against finite-element models, which we adapt for a viewing window on the tube.

The cryogenic sloshing work in Numerical Study on the Sloshing and Thermodynamic Characteristics of Liquid Hydrogen Storage Tank in Hydrogen-Powered Aircraft has the same passive-plus-disturbance structure (baffles plus sloshing) and explicitly lacks experimental validation. We treat it as a later target, not part of this study, because cryogenic rigs are far costlier than a heated water tube.

Research Program

The physical rig replicates the Yang et al. configuration: a stainless tube with a resin twisted belt, a constant-flux electric heating jacket, and Re 5,000–18,000. We add instrumentation that turns scalar measurements into fields. This means an axial and azimuthal grid of roughly 24 wall thermocouples, differential pressure taps, inlet and outlet bulk temperatures for an energy balance, and one short section where the jacket is replaced by a thin-film heater so the outer wall can be imaged in the midwave infrared, following Pauzauskie et al.

The actuation space is widened beyond the original study. It covers the three transducer types (21, 25, 28 kHz), several drive powers, and transducer counts of one to three placed at varied axial and azimuthal positions. The target is on the order of 150–250 steady-state runs, each a (Re, heat flux, actuation) triple, logged together with passive-only runs.

The passive surrogate is trained on RANS simulations of the insert-only tube across the Re and heat-flux range. It is then lightly calibrated on the passive rig runs only, so that any remaining error in the active runs can be attributed to actuation.

The hypernetwork encodes transducer placements as a set, so its output does not depend on the order in which transducers are listed. It concatenates this encoding with frequency, power and Re, and emits the singular values $s_\ell(\mu)$ for the modulated layers. It also emits a scalar correction for the friction head. Training minimizes

$$\mathcal{L} = \sum_{k}\big(T_k^{\text{pred}} - T_k^{\text{meas}}\big)^2 + \lambda_1 \big(\log f^{\text{pred}} - \log f^{\text{meas}}\big)^2 + \lambda_2\, \mathcal{R}_{\text{energy}},$$

where $k$ indexes thermocouple and infrared pixels, $f$ is the friction factor, and $\mathcal{R}_{\text{energy}}$ penalizes mismatch between applied heat and measured bulk enthalpy rise. The energy term addresses the known tendency of generated weights to give unphysical outputs outside the training range. Nusselt number and PEC are computed from predicted wall temperatures and bulk temperatures exactly as the experiment computes them, so model and rig are scored on identical quantities.

Baselines

The naive reference is the passive surrogate with no correction. It predicts the insert-only tube regardless of actuation, and its error on active runs measures how much ultrasound actually changes the field.

The strongest classical competitor is the engineer's standard calibration: a fitted enhancement factor. The simulated wall heat transfer coefficient is multiplied by a scalar $g(\mu)$, regressed from the same rig runs by a Gaussian process over frequency, power, transducer count and Re. A second strong competitor is a Gaussian-process discrepancy model on the full field. It fits the residual between the passive surrogate and each thermocouple as a function of sensor location and $\mu$, with transducer positions entering as distance features. These two represent what practitioners would actually do with a few hundred measurements.

The strongest neural competitor is hpFNO-style conditioning from Extending Fourier Neural Operators for Modeling Parameterized and Coupled PDEs, with bias shifts generated from $\mu$, together with plain input concatenation of $\mu$ to the primary network. Both are fine-tuned on the same rig data, starting from the same simulation-trained weights.

The ablations isolate the key ingredient. Rank zero with only a global scale is the scalar-enhancement baseline expressed inside the network. Modulation of scale and shift only follows Liu et al. and Wen et al. We then vary rank $r$ from 1 to 8, drop the frozen simulation prior, and drop the energy penalty.

Fairness is enforced by giving every method identical training runs, identical held-out splits, and the same sensor inputs. Every method gets an equal hyperparameter search budget, measured in trials, chosen on a validation subset of training layouts. Reporting uses five seeds or resamples.

No reported result was found on Lacuna for any predictive model of combined ultrasonic and twisted-insert heat transfer. The only reference numbers are physical: Nusselt gains of up to 19.06% (single 21 kHz transducer) and 28.06% (two transducers), and PEC of 1.15–1.36, from Yang et al. The closest surrogate result, hpFNO's 55–72% error reduction, is on synthetic PDEs and does not transfer as a target. Beating the baselines therefore means lower held-out field error and lower Nusselt/PEC error than both Gaussian-process calibrations. It also means correct ranking of held-out configurations, including reproducing that two transducers beat three at matched power, where the scalar factor gets the ranking wrong or with the wrong sign.

Why not just…?

Why not just fit a Gaussian process to Nusselt number versus the actuation settings? This is the most serious threat. We expect it to be competitive, possibly better, for interpolating scalar Nusselt number at settings that are densely sampled in frequency and count. Its weakness should appear in two places: spatial fields, and layouts whose transducer positions differ from those in training. There a scalar regressor has no notion of where along the tube the enhancement occurs. The leave-layout-out split tests this directly. If the Gaussian process matches our model there too, we will report that the hypernetwork adds nothing for this device, and narrow the claim to field prediction or drop it.

Why not just calibrate the simulation with a fitted enhancement factor on the wall heat transfer coefficient? This is the rank-zero member of our own model family, so the rank sweep answers it. If rank one or a global scale already reaches the error floor, the finding is that ultrasound acts as a nearly uniform wall enhancement, and the added complexity is unwarranted. Only a clear gain at rank above one justifies the method.

Why not just append the actuation settings as extra inputs and fine-tune? hpFNO reports that input augmentation steers parameterized operators poorly. However, that evidence comes from synthetic data with abundant samples. With a few hundred runs, full fine-tuning may overfit, or may do fine. The concatenation baseline, run from the same initialization and budget, settles this for the small-data regime.

Why not simulate cavitation directly and skip the learned correction? Cavitation under acoustic forcing in a swirling turbulent flow requires resolving bubble dynamics at timescales set by 21–28 kHz within a turbulent pipe flow. The file documents that even simpler multiphase transients are too costly for design loops and remain unvalidated (Numerical Study on the Sloshing and Thermodynamic Characteristics of Liquid Hydrogen Storage Tank in Hydrogen-Powered Aircraft). Neural surrogates also run orders of magnitude faster than 3D solvers (PDEBench reports about 0.14 s versus about 60 s). We do not test this alternative ourselves, and we say so.

Proposed Experiments and Decisive Checks

We use three held-out splits. Leave-one-frequency-out trains on two transducer types and tests the third. Leave-layout-out holds out entire transducer placement sets, including some three-transducer arrangements, while others remain in training. Leave-Re-band-out holds out the lowest and highest Reynolds bands. Metrics are relative $L_2$ error on thermocouple and infrared fields, absolute percentage error on Nusselt number, friction factor and PEC, Spearman rank correlation across held-out configurations, and energy-balance violation. A data-efficiency curve retrains all methods on 25%, 50% and 100% of the runs.

The first decisive check is the rank sweep compared against the two Gaussian-process calibrations on leave-layout-out. The core claim stands only if some rank above one gives clearly lower field and PEC error than both. If not, we conclude that ultrasonic enhancement is well described by a scalar factor.

The second check is prospective. We use each model to choose the actuation configuration it predicts maximizes PEC among untested layouts, build those configurations on the rig, and measure them. The proposal succeeds if the hypernetwork's chosen configuration measures at or above the best configuration in the training set, and above the Gaussian-process picks. It fails if its predicted optimum does not reproduce physically.

The third check is diagnostic. We compare the Fourier and U-Net primary networks near the transducer footprints, to test whether RecFNO's gradient-smoothing failure recurs in a convective device and whether a convolutional primary avoids it.