Learned Residual Feedback for Fast Re-Characterization of Radiation-Damaged Detectors

Semiconductor radiation detectors contain microscopic defects called charge traps. Traps capture electrons and holes as they drift toward the electrodes, so the detector reads out less charge than was deposited and its energy measurements blur. Radiation adds traps over time. Recent physics-based machine learning methods can map where the traps are inside a crystal. They build a differentiable simulator of charge transport, treat each region's mobility and trapping lifetime as trainable weights, and fit those weights to the measured electrode signals by gradient descent. The cost is the obstacle. In the simplest one-dimensional case, a fit takes about three hours on a GPU for a single crystal (Characterization of Crystal Properties and Defects in CdZnTe Radiation Detectors). The 3D voxel version also suffers from fading gradients and oscillating losses (A Learning-Based Physical Model of Charge Transport in Room-Temperature Semiconductor Detectors). That is tolerable for a one-time factory measurement. It rules out re-mapping a detector repeatedly as it accumulates radiation damage in a space instrument, a beamline, or a medical scanner.

Computer vision had the same bottleneck when every 3D scene was fit from scratch. It resolved it partly with feed-forward reconstruction networks and partly with learned refinement loops. In those loops a network looks at the current reconstruction's rendering error and proposes a correction, with no gradient descent at test time (ReSplat: Learning Recurrent Gaussian Splats). We propose the detector analogue, aimed at tracking damage over time. A small learned module repeats four steps. It simulates the electrode waveforms implied by the current trap map, compares them with the measured waveforms, and routes each waveform's error back to the voxels the simulated charge actually drifted through. It then updates those voxels' parameters. When a new radiation dose arrives, the module starts from the previous map and needs only a few such steps.

Core Claim

The claim is that a few iterations of simulator-in-the-loop, gradient-free refinement can recover 3D maps of drift and trapping parameters as accurately as per-crystal backpropagation, in seconds rather than hours. We also expect it to be more accurate than backpropagation in the voxels where gradients fade. Two ingredients make this plausible, and both are new relative to the file.

The first is how waveform errors are routed back into the crystal. In ReSplat, image-space errors reach 3D primitives through camera geometry and global attention. Here we use transport physics instead. The forward simulator already records, for every injected event, how much charge occupied each voxel at each time step. The Shockley–Ramo weighting potential, which describes how much a moving charge in a given voxel induces on each electrode, says how strongly that voxel contributes to each electrode's signal. We use these two quantities as fixed attention weights, so each voxel gathers only the waveform error segments it could have caused.

The second is the deployment regime. The detector is a sequence of states under increasing radiation dose, and each re-characterization warm-starts from the last. This is the detector counterpart of moving from static to dynamic reconstruction, as in DGS-LRM: Real-Time Deformable 3D Gaussian Reconstruction From Monocular Videos.

Formally, let $\theta$ be the voxel field of physical parameters: electron and hole drift coefficients, trapping and detrapping lifetimes, and recombination rate. Let $F(\theta)$ be the transport simulator's predicted electrode waveforms, and $s$ the measured waveforms. The refiner $U_\phi$, with learned weights $\phi$, iterates

$$\theta_{t+1} = \theta_t + U_\phi\big(\theta_t,\; B_{\theta_t}[\,s - F(\theta_t)\,]\big), \qquad t = 0,\dots,T-1,$$

where $B_{\theta_t}$ is the physics back-projection. It maps waveform residuals to per-voxel feature vectors using the occupancy and weighting-potential maps produced by the same forward pass. We train $\phi$ on simulated crystals by minimizing $\sum_t \lVert \theta_t - \theta^\star \rVert$, the error against ground-truth parameters at every iteration.

At test time this requires only forward simulations, with no backpropagation through the time-stepped transport. That removes the fading-gradient problem at its source. It does not remove genuine non-identifiability. For that reason the update head also outputs a per-voxel confidence, which we score for calibration.

Relation to Prior Work

The detector papers motivate the problem and supply the forward model. Ballester et al. fit 7 parameters in each of 100 depth layers with Adam. They report 0.43% average RMSE noise-free and about 3 hours per crystal, and they flag both the "inverse crime" (using the same simulator to generate and invert signals) and the missing lateral diffusion and repulsion (Characterization of Crystal Properties and Defects in CdZnTe Radiation Detectors). Banerjee et al. extend this to 3D sub-pixel voxels under an 11×11 pixel anode, with parameters as recurrent-network weights trained by backpropagation through time. They report a mean error of 0.0333 but poor convergence of electron recombination (A Learning-Based Physical Model of Charge Transport in Room-Temperature Semiconductor Detectors; A physics based machine learning model to characterize room temperature semiconductor detectors in 3D). Two further papers show, respectively, that electrode-only fitting needs total-variation smoothing to avoid erratic solutions, and that recombination remains hard to recover (RNMSE 0.3675) (Learning-based physical models of room-temperature semiconductor detectors with reduced data; Identifying Defects without a priori Knowledge in a Room-Temperature Semiconductor Detector Using Physics Inspired Machine Learning Model).

All four solve each crystal from scratch and carry nothing across crystals or across time. We keep their simulator and parameterization, including the "equivalent single trapping center" simplification, and replace the per-crystal optimizer with a learned, amortized refiner.

The vision papers supply the method template and also bound it. GS-LRM and DepthSplat show that one forward pass can replace per-scene fitting, at 0.23 s and 0.6 s respectively (GS-LRM: Large Reconstruction Model for 3D Gaussian Splatting; DepthSplat: Connecting Gaussian Splatting and Depth). They also show its limits: single-pass predictions blur fine detail and degrade under input shift. ReSplat's 4-step error-feedback loop, which beats DepthSplat by 3.5 dB PSNR on 8-view DL3DV, is the mechanism we transfer. We change two things. Our error is a physical waveform residual instead of an image error, and our routing comes from transport trajectories instead of camera projection.

The classical signal-processing corrections set the practical bar we must eventually clear. They model trapping with one bulk time constant and gain about 30% in energy resolution (Determining the drift time of charge carriers in p-type point-contact HPGe detectors; Charge trapping correction and energy performance of the Majorana Demonstrator). Martin et al. note that their drift-time correction fails in the outer 35% of the detector volume.

Research Program

We first rebuild the two synthetic settings in the file, since the file does not indicate that either is released. The first is the 1D 100-layer stratified CdZnTe (CZT) setup. The second is the 3D 11×11-pixel crystal at 100 µm voxel resolution, with drift, trapping, detrapping and recombination.

To create training populations, we draw parameter fields from a structured prior that mixes three components: smooth bulk variation, localized inclusion-like pockets, and damage fields. The damage fields grow trap density with dose along a simulated beam profile. The file gives no quantitative CZT dose-to-trap rate. We therefore treat the damage law as a family to sweep, and we do not claim it is calibrated.

A second, higher-fidelity simulator generates held-out test signals. It adds lateral charge-cloud diffusion and Coulomb repulsion. The refiner never runs this simulator inside its loop, because the in-loop simulator stays the simplified one.

The refiner itself has three parts:

  • an initializer that predicts $\theta_0$ from the raw waveforms, or takes the previous dose step's estimate;
  • the physics back-projection $B$;
  • a sparse 3D convolutional update head shared across iterations, with $T$ between 4 and 8.

Analysis then focuses on three questions:

  • where the refiner beats backpropagation, broken down by per-voxel charge occupancy;
  • whether the learned prior removes the need for total-variation regularization;
  • how much spatial granularity matters for downstream energy resolution versus an equivalent bulk model. The file leaves this tension open.

Baselines

The naive reference is a bulk single-time-constant correction of the kind used by Martin et al. and the MAJORANA collaboration (Determining the drift time of charge carriers in p-type point-contact HPGe detectors; Charge trapping correction and energy performance of the Majorana Demonstrator). We extend it with the spatially uniform equivalent-trap fit of Learning-based physical models of room-temperature semiconductor detectors with reduced data. It is fit on the same waveforms and compared on downstream photopeak width (FWHM). It is the "do nothing spatially" point. If it matches us on energy resolution, the spatial map is not worth its cost for that use.

The strongest existing method is per-crystal backpropagation through time on the voxel recurrent model of A Learning-Based Physical Model of Charge Transport in Room-Temperature Semiconductor Detectors. It uses that paper's Adam schedule with learning rate stepped from 5e-4 to 1e-5, plus total-variation regularization. That paper reports a mean parameter error of 0.0333, with electron drift error 0.0117 and hole drift error 0.0152. In 1D the corresponding reference is Ballester et al.'s 0.43% noise-free RMSE at about three hours per crystal. We found no reported 3D wall-clock result on Lacuna. We will rerun this baseline on our regenerated data, using the same simulator, noise levels, and GPU, and report error-versus-wall-clock curves rather than single points. We count a win if the refiner matches its final error (within run-to-run spread) at at least 100× lower wall-clock time. We count a stronger win if it achieves lower error on low-occupancy voxels and on the mismatched-simulator test.

Three ablations isolate the key ingredient. The first is a one-shot amortized regressor, which is our initializer alone with $T=0$. The second replaces the residual input with zeros, so the head refines without feedback. The third replaces the physics back-projection with unstructured global attention, as in ReSplat. A fourth variant feeds the update head the backpropagated gradient instead of the routed residual. This is a learned-optimizer baseline, and it tests whether avoiding gradients matters in itself. All learned variants get the same parameter count, training set, and training compute.

Why not just…?

Why not just use bulk signal-processing correction? It is cheap and gains about 30% in resolution in point-contact germanium (Charge trapping correction and energy performance of the Majorana Demonstrator). However, it already fails in the outer 35% of the volume in Determining the drift time of charge carriers in p-type point-contact HPGe detectors. It also cannot represent localized, dose-dependent trap pockets under a pixelated anode. This is a real threat for spatially mild damage. The energy-resolution comparison across damage-field types will show where, if anywhere, the spatial map pays off, and we will report the regimes where it does not.

Why not just run per-crystal Adam with more compute or a better schedule? Extra compute does not fix gradients that fade in voxels the charge barely visits. That fading is the documented cause of poor recombination recovery (A physics based machine learning model to characterize room temperature semiconductor detectors in 3D). Extra compute also multiplies the hours per re-characterization across a damage sequence. The occupancy-binned error analysis tests the first point directly, and the time-to-tracking-error comparison tests the second.

Why not just predict the map in one pass and polish it with a few Adam steps? This warm-start hybrid is the most serious competitor, and it may well suffice in-distribution. We run it under equal wall-clock budgets. The refiner has to win in two places to justify itself. The first is the mismatched-simulator test, where the in-loop simulator is wrong and Adam will chase the mismatch. The second is low-occupancy voxels. If the hybrid matches the refiner in both, we will report that the residual-feedback loop adds nothing beyond amortized initialization.

Proposed Experiments

The first check is 1D sanity at matched accuracy. On the 100-layer setup at noise-free, 0.1%, and 2% noise levels, we compare refiner error and runtime against per-crystal Adam.

The second is the 3D headline. On the 11×11-pixel crystal, we report the per-parameter error metric of Banerjee et al. for every baseline and ablation. We break it down by voxel charge occupancy and pay particular attention to electron recombination.

The third is the inverse-crime test. We invert signals from the higher-fidelity diffusion-plus-repulsion simulator using the simplified in-loop simulator. This is the decisive check that refinement driven by physical residuals degrades more gracefully than one-shot regression or gradient fitting.

The fourth is damage tracking. Over sequences of 10–20 dose steps, we measure parameter error per step, the number of refinement steps needed, and cumulative wall-clock time. We compare warm-started refinement against from-scratch and warm-started Adam.

The fifth is the downstream test. Using each method's map, we correct simulated monoenergetic photopeaks event by event and report FWHM against the bulk single-time-constant correction.

We also score how well the confidence head is calibrated: per-voxel confidence against actual error. This shows whether the model knows when it is guessing in regions with no signal.

Validation on measured CZT waveforms is the necessary next step, but the file contains no such dataset. All claims above are therefore restricted to simulation, with the mismatch test as our best guard against overfitting to one simulator.