High-power accelerator targets are run close to the edge. An isotope-production target must stay above 2000°C so that rare isotopes diffuse out fast enough, yet only a narrow margin separates that temperature from sintering or melting. An astrophysics target must survive tens of kilowatts per square centimetre. Operators know the design in general, but not the particular disk now in the beam. Its emissivity, porosity, contact resistance and beam alignment all differ a little from nominal, and a margin set for the nominal disk can be wrong for this one. Data on this disk cannot be gathered freely, because the only safe measurements come from the first low-current minutes of a ramp-up. We propose to borrow the answer the speech community found for an analogous problem. Speech recognition and synthesis models must adapt to one new speaker from seconds of audio, and the methods that work in that regime update a handful of physically meaningful parameters rather than generic weights.
Concretely, we would train a fast neural stand-in for a finite-element thermal simulation, which we call a surrogate, on many simulated target variants. When a new physical target goes into the beam, we adapt the surrogate using a few pyrometer and thermocouple readings from low-current steps. The adaptation updates mainly a small set of physical knobs: per-disk emissivity, interface conductance and beam-spot width. A small, heavily constrained correction term absorbs what the knobs cannot. The adapted model is then used to predict the hottest point at the full planned current before the operator raises the current. The question is whether this speech-derived recipe predicts those temperatures more accurately than generic adaptation, plain fine-tuning, or classical calibration of the simulator. A second question is whether separating "what this target is" from "how it is degrading" lets the model flag sintering drift early.
Core Claim
Three results from low-resource speaker adaptation make the analogy specific rather than decorative.
The first is that physically parameterized adaptation wins at the smallest data budgets. In Acoustic Model Adaptation from Raw Waveforms with Sincnet, the first layer of a raw-waveform recogniser is constrained to bandpass filters described only by their cut-off frequencies. Updating just those 80 numbers acts like vocal-tract-length normalisation and cuts word error rate on children's speech from 68.19% to 31.65%.
The second is that staged, subspace-structured adaptation beats generic linear transforms. A Two-stage Speaker Adaptation Approach for Subspace Gaussian Mixture Model based Nonnative Speech Recognition first fits shared covariance and then refines the speaker projections. It reaches 13.99% WER where the generic linear feature transform reaches 38.83%.
The third is that decoupling stable identity from a separate deficiency factor helps. Structured Speaker-Deficiency Adaptation of Foundation Models for Dysarthric and Elderly Speech Recognition gives speaker identity and severity of speech impairment separate small adapter modules, and reports 19.45% WER on UASpeech.
These results also carry negative lessons that constrain the design. Generic bottleneck and convolutional adapters still trail full fine-tuning below an hour of data (Investigating Adapters for Parameter-efficient Low-resource Automatic Speech Recognition). Adaptation that conditions layer-normalisation scale and bias on a speaker embedding degrades sharply below about ten sentences (AdaSpeech: Adaptive Text to Speech for Custom Voice). Full fine-tuning on scraps of data memorises noise (Voice Filter: Few-shot text-to-speech speaker adaptation using voice conversion as a post-processing module).
The claim follows from all three. Per-target thermal prediction from a few low-current measurements is the same structural problem: a large pretrained model, one new instance, and data that is scarce for a hard physical reason. In speech the reason is that dysarthric speakers tire. Here it is that one cannot probe a target near failure. In this regime, adaptation restricted to physical parameters, done in two stages and with static and drifting factors separated, should extrapolate to full current better than generic adapters, fine-tuning, or curve fitting.
The model we adapt is $$\hat T(x) = f_\theta(x;\phi) + g_\psi(x),$$ where $x$ is the operating condition and probe location (beam current, energy, time), $f_\theta$ is the pretrained surrogate with frozen weights $\theta$, and $\phi$ is the small vector of physical knobs. The term $g_\psi$ is a low-capacity residual with parameters $\psi$, penalised toward zero, which absorbs model-form error the knobs cannot express. The physics explains why so few knobs can matter so much. A radiatively cooled disk stack obeys $P \propto \varepsilon A T^4$ (power radiated scales with emissivity $\varepsilon$, radiating area $A$ and the fourth power of temperature, as exploited by the SPES multi-disk design). An emissivity error therefore shifts the equilibrium temperature as $\varepsilon^{-1/4}$ at every current. This is a low-dimensional, structured deviation, much as vocal-tract length warps every frequency band coherently.
Research Program
Regimes and simulated library. The work covers two thermal regimes chosen to stress different physics. The radiation-dominated regime follows the SPES isotope-separation-on-line target: seven uranium-carbide disks, a 35–70 MeV proton beam at 0.2–0.7 mA, and 8–10 kW deposited (The SPES radioactive ion beam project of LNL: status and perspectives; The SPES project of INFN: Facility and detectors; design details in Pretea et al. and de Angelis et al.). The conduction-dominated regime follows the JUNA rotating solid target, which is rated for 20 kW/cm² under a 12 mA proton beam (Progress of Jinping Underground Nuclear Astrophysics Experiment (JUNA) project). For each regime we build a parametric finite-element model and sample thousands of variants over emissivity, porosity-dependent conductivity, interface conductance, disk spacing, beam width and offset, and, for JUNA, rotation speed. The surrogate is trained on the resulting temperature fields.
Ground truth for new targets. "New targets" are generated by a higher-fidelity simulator than the one used for training, so that adaptation faces genuine model-form error and not only parameter error. The higher-fidelity simulator includes disk warping that changes the view factors between disks, temperature-dependent emissivity, and conductivity that falls as sintering progresses.
Adaptation protocol. Adaptation follows the two-stage recipe. Stage one fits the knobs shared across disks, such as global emissivity scale and beam width, from the first one or two current steps. Stage two refines per-disk knobs and the residual. Following the structured speaker-deficiency idea, the knob vector is split into a static part, fitted once, and a drift part that evolves under a slow-change prior. The drift part's trajectory then serves as a degradation indicator.
Measurement budget. Measurements mimic what a real facility has: a pyrometer view of the front disk and a few thermocouples on the containment box, at three to five low-current steps, for example 0.05–0.2 mA in SPES.
Real data. Real-target validation would need facility collaboration and is a stated dependency. The simulated benchmark is designed to be released regardless.
Baselines
The naive reference is the zero-shot surrogate run with nominal design parameters and no adaptation, together with the facility-style practice of applying a fixed safety margin to it. This is what "do nothing" means, and it sets the margin any method must shrink.
The most serious competitor is classical calibration of the finite-element simulator itself. Here the same physical knobs are fitted by least squares or Bayesian inversion to the same few measurements, then the full simulator is run at target current. It is the strongest competitor because it uses identical physics and identical data, and it lacks only speed and a way to absorb model-form error. It will receive the same measurements and the same prior ranges, with a wall-clock budget matched to a realistic ramp-up pause and also an unlimited budget. To beat it, our method must match its peak-temperature error at full current within the matched time budget. Under model-form error it must do strictly better, through the residual term.
The two strongest learning-based competitors come from speech. The first is generic residual adapters on the frozen surrogate, both feed-forward bottleneck and one-dimensional convolutional variants, as in Investigating Adapters for Parameter-efficient Low-resource Automatic Speech Recognition. The second is scale-and-bias conditioning on a learned per-target embedding, as in AdaSpeech, together with full fine-tuning under weight decay and early stopping. All receive the same measurements, the same frozen backbone and the same hyperparameter search budget, with selection by leave-one-current-step-out validation.
There is no reported result found on Lacuna for per-target thermal prediction accuracy by any of these methods. The quantitative anchor from the cited papers is therefore the speech-side gap. Physical-parameter adaptation took children's speech from 68.19% to 31.65% WER with 80 parameters (SincNet), and subspace adaptation reached 13.99% against the generic linear transform's 38.83% (two-stage SGMM). We will count the proposal as successful if it reduces full-current peak-temperature error by a clear margin over the best generic adapter at three to five measurements.
Three ablations isolate the key ingredient. The first removes the residual term, leaving physical knobs only. The second replaces the physical knobs with an equal number of unconstrained latent parameters, which tests whether physical meaning, rather than parameter count, does the work. The third fits jointly instead of in two stages, and merges the static and drift parts into one vector.
Why not just…?
Why not just calibrate the simulator directly and skip the neural surrogate? This is the real threat, and we treat it as a co-equal hypothesis rather than a strawman. If direct calibration matches our error at every budget and runs within a ramp-up pause, the learned surrogate adds nothing and we will report that. Our expectation that it falls short rests on two points. First, a calibrated simulator cannot represent physics it omits, such as warping-induced view-factor changes, so it carries systematic error at extrapolated currents. Second, Bayesian inversion of a full finite-element model under a tight time budget yields poor uncertainty estimates. The model-form-error experiment decides this.
Why not just fine-tune the surrogate, or attach standard adapters? Speech evidence says both struggle at tiny budgets. Full fine-tuning memorises measurement artefacts (Voice Filter). Bottleneck and convolutional adapters trail full tuning below an hour of audio (Investigating Adapters). With three to five noisy temperature readings, a generic adapter has no reason to extrapolate along the $T^4$ structure, whereas emissivity does. The latent-knob ablation shows whether physical meaning or mere smallness matters.
Why not just fit a curve or Gaussian process to the low-current readings and extrapolate? A data-only regressor treats the jump from 0.2 to 0.7 mA as unconstrained extrapolation. It cannot know that the front-disk temperature should scale through a radiative balance or that conduction paths saturate. We include a Gaussian process with a physically motivated mean, temperature proportional to the fourth root of power, as a baseline. If it matches us, the surrogate's spatial structure is unnecessary for peak prediction, and the contribution narrows to locating the hottest point in disks the pyrometer cannot see.
Why not just keep a generous fixed safety margin? This costs beam time and yield. The SPES design exists precisely because isotope release rates demand running near the material limit (The SPES Radioactive Ion Beam facility of INFN). We measure how much allowable current each method recovers at equal risk of exceeding the limit.
Proposed Experiments
The primary metric is the absolute error in predicted peak temperature, in kelvin, at the full planned current, across held-out higher-fidelity targets, as a function of the number of low-current measurements (one to eight). The secondary metrics address the operational decision. One is the coverage of the predicted 95% interval. The other is the "safe current" each method would authorise without the true temperature crossing a sintering threshold, scored by recovered current at a fixed violation rate.
The first decisive check is the data-budget curve. The speech literature predicts that physical-knob adaptation dominates at the smallest budgets and that generic adapters catch up as data grows. Seeing that crossover in thermal data would confirm the analogy is mechanistic. Seeing no crossover would refute it.
The second check is the model-form-error sweep. We increase warping and temperature-dependent emissivity in the ground-truth simulator. Classical calibration should degrade faster than knobs plus residual, and if it does not, the residual term is not earning its place.
The third check is drift detection. On synthetic runs with progressive sintering, we test whether the separated drift parameters move before the peak-temperature error grows, compared with monitoring raw pyrometer residuals.
The fourth check is cross-regime transfer. Repeating everything in the conduction-dominated JUNA regime tests whether the recipe depends on radiative physics or holds generally.
Finally, a stress case adds beam-profile fluctuations, the failure mode the SPES literature names. It checks that beam width, as a knob, absorbs them rather than leaking them into emissivity.
References
- Acoustic Model Adaptation from Raw Waveforms with SincNet
- A Two-stage Speaker Adaptation Approach for Subspace Gaussian Mixture Model based Nonnative Speech Recognition
- Structured Speaker-Deficiency Adaptation of Foundation Models for Dysarthric and Elderly Speech Recognition
- Investigating Adapters for Parameter-efficient Low-resource Automatic Speech Recognition
- AdaSpeech: Adaptive Text to Speech for Custom Voice
- Voice Filter: Few-Shot Text-to-Speech Speaker Adaptation Using Voice Conversion as a Post-Processing Module
- The SPES radioactive ion beam project of LNL: status and perspectives
- The SPES project of INFN: Facility and detectors
- The SPES project at the INFN- Laboratori Nazionali di Legnaro
- Progress of Jinping Underground Nuclear Astrophysics Experiment (JUNA) project
- The SPES Radioactive Ion Beam facility of INFN