Big particle detectors now produce more data than their cables can carry, so they compress it on the detector with small neural networks burned into radiation-hard chips. These compressors are trained to reproduce their input faithfully on average. In precision physics, though, the real enemy is no longer random noise. It is small, structured biases called systematic uncertainties, such as an energy scale that is off by half a percent or a background shape that a selection cut has quietly distorted. A compressor that is accurate on average can still turn a calibration drift into a distortion that no later software can undo. This note proposes training and searching for on-detector compressors under a property we call calibration transparency. Applying a calibration change and then compressing should give the same result as compressing and then applying the change. If that holds, the standard offline corrections still work on compressed data, and the compressor adds no new systematic floor.
The hardware side of the question is shaped by heat. The limit is fixed by front-end power ceilings of about 100 mW per chip, and by the fact that any logic stacked above existing circuits must be fabricated below roughly 400°C. We treat this as the budget that sets how much model the detector can afford. We then ask how much of that budget calibration transparency costs on a planar ASIC, on a commercial FPGA, and on a projected monolithic 3D FPGA built from low-temperature oxide transistors.
Core Claim
The closest existing system is the CMS High-Granularity Calorimeter compression chip of A reconfigurable neural network ASIC for detector front-end data compression at the HL-LHC. It runs a quantized encoder on the detector in 25 ns at 95 mW and scores fidelity with Earth Mover's Distance, a measure of how much energy must be moved to turn the reconstructed shower into the true one. That metric is an average and says nothing about how the compressor responds to nuisance parameters. A nuisance parameter is any detector property not known exactly, such as gain, noise level, or energy scale. Our claim is that the systematic damage a compressor can do is captured by its failure to commute with these known detector transformations. We further claim that this failure can be trained down cheaply, because the penalty acts only during training and adds nothing to the circuit that runs inference. Concretely, let $x$ be raw sensor data, $T_\theta$ the detector transformation for nuisance value $\theta$ (for example, an energy-scale shift), $E$ and $D$ the encoder and the off-detector decoder, and $\Theta$ the calibration envelope the experiment expects. We train
$$\mathcal{L} = \mathcal{L}_{\text{rec}}(D(E(x)),x) + \lambda\, \mathbb{E}_{\theta\sim\Theta}\big\| D(E(T_\theta x)) - T_\theta\, D(E(x)) \big\|^2,$$
with quantization applied inside the training loop, so that the property survives at the bit widths the chip actually uses.
This differs from the adversarial decorrelation of Decorrelated jet substructure tagging using adversarial neural networks. That work makes a classifier's output independent of jet mass. It operates offline at full precision and targets one observable. Our method targets a representation, not a decision. It asks for commutation, which preserves the nuisance dependence so that calibration can remove it, rather than asking for independence. It is evaluated under hardware quantization. It also differs from the hardware codesign search of Neural architecture codesign for fast physics applications, which trades accuracy against bit operations and never measures systematics. We keep that search machinery but change what it optimizes.
Research Program
The first stage is a measurement study, and it motivates everything else. We take a reconstruction-trained, quantized encoder in the style of the HGCAL ASIC and apply energy-scale, gain, and noise shifts drawn from a realistic calibration envelope. We then measure how far compression-then-calibration departs from calibration-then-compression. We propagate that departure into a downstream measurement and report two quantities: the shift in a reconstructed resonance mass and the profiled discovery significance. Profiled significance is significance computed with nuisance parameters fitted alongside the signal. If commutation error is negligible at realistic bit widths, the proposal stops here with a useful negative result. That would be an assurance that existing compressors are systematically safe.
If the error is not negligible, the second stage trains with the objective above, sampling $\theta$ from the calibration envelope during training. The same study applies to an offline classifier, which addresses the open question flagged in the file of whether adversarial decorrelation survives 6–8 bit quantization and pruning. The third stage runs the codesign search of Neural architecture codesign for fast physics applications with profiled significance and commutation error added as objectives next to bit operations. This produces Pareto fronts, the curves of best achievable trade-offs, of physics reach against hardware cost.
The fourth stage maps those fronts onto three hardware cost models. The first is the planar 65 nm ASIC envelope, with 50 ns latency and a power ceiling under 100 mW. The second is a commercial FPGA. The third is a monolithic 3D FPGA modeled in Verilog-to-Routing using the reported cost ratios of Monolithic 3D FPGAs Utilizing Back-End-of-Line Configuration Memories: 3.4× lower area-time-squared and 60% lower configuration static power. That design is feasible only because its oxide transistors are processed below 400°C, as discussed in The Roadmap of 2D Materials and Devices Toward Chips. The same constraint saddles it with slow p-type devices, which we include as a latency penalty. This stage is a projection, not a fabrication. The question it answers is whether reconfigurable stacked logic, which can be reloaded with recalibrated weights, shifts the systematics-versus-power front enough to matter.
Baselines
The naive reference is the reconstruction-only quantized autoencoder, the HGCAL ASIC recipe, trained with Earth Mover's Distance and no nuisance awareness. Its hardware figures are 50 ns total latency, 95 mW, and 2.4 nJ per inference, as reported in A reconfigurable neural network ASIC for detector front-end data compression at the HL-LHC. We found no reported Earth Mover's Distance value on Lacuna for that chip, so we retrain it ourselves on identical simulated showers with the same bit widths, architecture family, and tuning budget. Beating it means matching its reconstruction fidelity and hardware cost while reducing commutation error and recovering profiled significance under the calibration envelope.
The second baseline is the uncompressed reference: full-precision raw data fed to the downstream analysis. It is the ceiling that no compressor can exceed. We report what fraction of its profiled significance each compressor retains.
The strongest existing method for the systematic question is downstream adversarial decorrelation. Decorrelated jet substructure tagging using adversarial neural networks reports 3.67σ for its decorrelated tagger versus 2.82σ for a conventional one under a 50% background uncertainty. We reproduce this setup on Delphes Z′ events in two arrangements: placing the decorrelation offline, after a naive compressor, and applying it to a quantized on-chip classifier. Both get the same search budget as our method. If offline decorrelation after a naive compressor matches our significance, our compressor-side penalty is not needed.
The hardware-search baseline is the codesign pipeline of Neural architecture codesign for fast physics applications. Its Tiny jet model reaches 70 ns on a VU13P FPGA, with a 30.25× reduction in bit operations at a 2.8% accuracy cost. We rerun it with the identical search budget, with and without our added objectives.
The key ablation separates the commutation penalty from simple training-time augmentation. We train on nuisance-shifted inputs with ordinary reconstruction loss, so the model sees the same variations but is not asked to commute with them.
Why not just…?
Why not just fix the bias offline, with standard calibration or a decorrelated downstream classifier? This is the most serious alternative. If a naive compressor merely shifts its outputs in a way the decoder can learn to invert, offline methods suffice, and the experiments are designed to detect exactly this. The reason to expect otherwise is that a nonlinear, low-bit encoder can respond to an energy-scale shift in a way that depends on the event: it rounds some showers across thresholds and leaves others alone. That information is gone before it reaches any server. The baseline comparison settles this directly. If offline decorrelation after a naive compressor recovers the significance, we will report that, and the contribution reduces to the measurement study.
Why not just train with nuisance-shifted data augmentation? Augmentation is cheaper and may give most of the benefit. That is why it is our central ablation and not a side note. Augmentation teaches the model to reconstruct shifted inputs well. It does not require the shift to pass through unchanged so that calibration can later remove it. A compressor trained this way could reconstruct shifted inputs accurately in Earth Mover's terms and still treat a drift as part of the signal. If the ablation matches the full objective, we adopt augmentation and say so.
Why not just spend more bits or a bigger model? Commutation error should fall as distortion falls, so more capacity is a legitimate fix if the hardware can pay for it. The codesign fronts answer this quantitatively: they show whether the penalty reaches a given systematic level at fewer bit operations than brute capacity does. The front-end power ceiling is the reason the gap matters.
Why not just retrain the chip's weights over its reconfiguration interface as conditions drift? The HGCAL ASIC can already reload its weights. But Solving Simulation Systematics in and with AI/ML notes that every retraining on finite calibration samples introduces second-order systematics that no one knows how to propagate. A calibration-transparent compressor would need retraining far less often. We test this by counting how wide a drift range each model tolerates before its downstream bias exceeds the statistical error.
Proposed Experiments
The decisive checks run on three settings, matching what the cited work used: simulated calorimeter showers in the HGCAL style, the Delphes boosted Z′ search, and the eight-particle Deep Sets jet benchmark. The primary metric is profiled discovery significance under a fixed calibration envelope. Secondary metrics are the commutation error, the shift in background mass shape after selection, and reconstruction Earth Mover's Distance. Hardware metrics are bit operations, latency and power synthesized through hls4ml, and Verilog-to-Routing area-time-squared for the stacked-FPGA projection.
The idea succeeds if three things hold together. First, a quantized commutation-trained encoder must recover a clearly larger share of the uncompressed reference's profiled significance than the naive encoder at equal hardware cost. Second, it must beat both offline decorrelation after naive compression and the augmentation-only ablation. Third, it must hold up across a wider drift range without retraining. The idea fails, and we report that plainly, if at 6–8 bits naive compressors already commute to within statistical error, or if augmentation alone reaches the same front. The stacked-FPGA projection is judged separately. It counts as interesting only if its power and area savings allow a systematically safe model inside the 100 mW envelope that the planar ASIC cannot fit.
References
- A reconfigurable neural network ASIC for detector front-end data compression at the HL-LHC
- Decorrelated jet substructure tagging using adversarial neural networks
- Neural architecture codesign for fast physics applications
- Monolithic 3D FPGAs Utilizing Back-End-of-Line Configuration Memories
- The Roadmap of 2D Materials and Devices Toward Chips
- Solving Simulation Systematics in and with AI/ML