Angle-resolved photoemission spectroscopy (ARPES) shines light on a crystal and counts the electrons that fly out at each energy and angle. Those counts form an image of the material's band structure, the map that sets how electrons move. Modern beamlines produce terabytes of these images during short, scarce sessions, and nobody can watch them all. The things worth catching are rare and not known in advance. They include a gap opening where none was expected, a band folding over because a symmetry broke, a surface degrading under the beam, or a detector artifact. This is the setting for reconstruction-based anomaly detection: train a model to rebuild only "normal" data, then flag inputs it rebuilds badly. That approach has a known weakness. If the model has too little capacity, it blurs real fine structure and raises false alarms. If it has too much, it learns to copy anything, anomalies included, and the alarm goes silent. No principled way of choosing the capacity without anomaly labels exists. Studies of how human vocabularies balance being simple against being informative point to a way out: stop picking one point on the trade-off and measure the whole curve.
We propose to score each spectrum by the number of bits a model trained on normal spectra must spend before its reconstruction is as good as the physics allows. ARPES is unusual among anomaly-detection domains because the noise is known. Electron counts are Poisson-distributed, so there is an absolute "good enough" level: a residual consistent with shot noise. A normal spectrum should reach that floor at a low bit cost. An anomalous one should need many more bits, or never reach it. This turns the capacity choice from a tuning knob into the measurement itself.
Core Claim
Train a variable-rate discrete autoencoder on normal ARPES spectra. Its code is built from stacked quantizer stages, so each additional stage adds a fixed number of bits and refines the reconstruction. For a spectrum $x$, let $D_r(x)$ be the mean Poisson deviance between the observed counts and the predicted intensity when only $r$ bits of code are used. Let $\tau(x)$ be the deviance that pure shot noise would produce around that predicted intensity, computed by Monte Carlo from the prediction itself. The anomaly score is
$$s(x) = \min \{\, r : D_r(x) \le \tau(x) \,\},$$
the smallest bit budget at which the reconstruction becomes statistically indistinguishable from noise. Spectra that never reach the floor get the maximum budget plus the remaining excess deviance. Localization comes from the per-pixel deviance at the rate where most normal spectra reach the floor.
The claim we will test is narrow. On ARPES anomalies, this score is more reliable than residuals at any single fixed capacity, and its advantage comes from two things together: using the whole rate curve, and having a physically calibrated stopping point. The claim is falsified if a fixed-capacity model with the same Poisson-normalized residual matches it.
Why the Trade-Off Literature Matters Here
The simplicity–informativeness literature supplies the lens rather than the method. Its central finding concerns what happens when each pressure acts alone. Carlsson et al. and Dubhashi et al. show in neural simulations of colour naming that a pressure toward easy learning alone collapses vocabularies into a few useless terms, while a pressure toward communication alone explodes them into hyper-specific ones. Only both together produce efficient systems. Zaslavsky et al. find that real naming systems sit near the optimal frontier of the information bottleneck, a formal trade-off between the bits a code uses and how much it preserves.
These two failures are exactly the over-smoothing and identity-copying failures of reconstruction detectors. The literature motivates our reframing: an input is anomalous if it is expensive to describe in the code that is efficient for normal data. It does not mean that inputs which need many descriptive bits are anomalous.
Tucker et al. supply the mechanism. Their vector-quantized variational information bottleneck (VQ-VIB) gives agents discrete symbols whose rate–accuracy balance is set by one coefficient. They also report that richer combinatorial codebooks can destabilize alignment with meaningful structure. That result constrains our design toward nested, staged codes, where rate rises one stage at a time, rather than one large combinatorial codebook.
On the anomaly side, Do Autoencoders Need a Bottleneck for Anomaly Detection? is the paper that most directly contradicts any "tighter bottleneck is better" story. It reports non-bottlenecked models beating bottlenecked ones on CIFAR versus SVHN (0.849 versus 0.696 AUROC), but losing by 0.1 AUROC on FashionMNIST versus MNIST. We read this as evidence that the right capacity depends on the data, which is the motivation for not fixing it. Video Anomaly Detection with Semantics-Aware Information Bottleneck applies a bottleneck regularizer at a single trade-off point for video. We instead read scores off the whole curve.
On the ARPES side, three papers supply methods but none detects anomalies. ARPESNet is a convolutional autoencoder trained "noise-to-noise" (on pairs of independent noisy measurements of the same spectrum) that compresses 256×256 slices to a fixed 512-dimensional code. It reports MSE 0.38 and PSNR 41.34 on its 19-material library. We reuse its architecture and training trick, but replace the fixed code with staged quantizers and the squared-error loss with Poisson deviance. Liu et al. remove detector-mesh grids by fitting twin networks to each spectrum. That is too slow for streaming use, and it targets one known artifact. Xian et al. reconstruct band surfaces with a probabilistic smoothness model warm-started from density functional theory calculations. That approach requires knowing the band count in advance and so cannot flag unexpected bands. Makino et al. is the closest materials precedent for reconstruction error as a physical anomaly score, but it works on composition vectors, not measurements.
Research Program
The model is an ARPESNet-style encoder and decoder with residual vector quantization in the middle. The first quantizer stage captures the coarse dispersion, and each later stage quantizes what the earlier ones missed. Training samples a random number of active stages per batch, so one network serves every rate. The decoder outputs a positive intensity map, and the loss is the Poisson negative log-likelihood of a second, independent noisy realization. This is the noise-to-noise idea expressed in the correct noise model.
The library provides paired realizations by splitting acquisition sweeps. Where raw detector counts are not available, we will estimate a per-dataset gain by matching variance to mean in flat regions. This is a real risk, and the failure of that calibration is itself something we will report.
The analysis has three parts. First, we will check that normal held-out spectra from training materials reach the shot-noise floor at a consistent, low rate. Second, we will characterize the rate curves of anomalous spectra and test whether different anomaly types leave different curve shapes. For example, a missing detector-grid correction should never reach the floor, while a small gap should reach it late. Third, we will inspect which code stages become active, testing the speculative idea from the trade-off literature that early stages behave like a reusable vocabulary of dispersion shapes. That last part is exploratory and not part of the main claim.
Baselines
The naive reference is the simplest residual score that needs no learning. A local polynomial or Gaussian-smoothed version of each spectrum serves as its "reconstruction," and the score is the Poisson deviance of the counts from it. If this matches the proposal, the learned normal manifold adds nothing.
The strongest existing reconstruction competitor is ARPESNet itself at its published 512-dimensional capacity, retrained on our splits. We score it in two ways: its native squared-error residual, and a Poisson-deviance residual. The second variant isolates how much of any gain comes from the noise model alone. We will also run a capacity sweep of separate fixed-capacity ARPESNets (16 to 2048 latent dimensions) and report both the best single capacity chosen in hindsight on the test anomalies, which is an oracle upper bound, and the mean of their normalized residuals.
From the non-bottleneck line, we will run an overcomplete autoencoder with skip connections, following Do Autoencoders Need a Bottleneck for Anomaly Detection?. From the feature-based line, we will use DINO features with k-nearest-neighbour distance to normal training spectra as the anomaly score. Ekahana et al. report that this feature extractor reaches 92.94% accuracy with 20 neighbours (versus 90.40% for supervised ResNet50) on Fe₃Sn₂ gapped versus gapless classification. Theirs is a supervised classification number, not an anomaly score. No reported result was found on Lacuna for anomaly detection on any ARPES dataset, so all anomaly numbers here will be new.
The key ablations are the proposal's own model scored at a single fixed rate, and the full curve scored with a learned threshold in place of the shot-noise floor. All methods share the same training spectra, the same GPU-hour budget, and the same validation protocol, which uses only normal data. No method sees anomalies when choosing hyperparameters, except the clearly labelled oracle. Beating the field means a higher image-level AUROC than both Poisson-normalized fixed-capacity ARPESNet and the capacity-sweep ensemble, with non-overlapping confidence intervals across seeds, on at least two of the three anomaly families described below. It also means a lower false-alarm rate on normal spectra from held-out materials.
Why not just…?
Why not just tune the latent size of ARPESNet? Tuning requires anomalies to validate against, and those are by definition unavailable at the beamline. The contradictory results in Do Autoencoders Need a Bottleneck for Anomaly Detection? suggest that the best size shifts with the data distribution. We include the oracle-tuned capacity as an upper bound. If our score does not approach that bound, the method fails on its own terms.
Why not just average residuals across several fixed-capacity autoencoders? This is the real threat, because a capacity ensemble also sees the whole curve. Our method differs in two ways: it uses one shared nested code instead of independently trained models, and it applies an absolute physical stopping rule instead of averaging. The capacity-sweep ensemble with Poisson normalization is run as a baseline. If it matches us, we will report that the contribution reduces to the noise-floor calibration, and that cheaper ensembles suffice.
Why not just use the Poisson-normalized residual at one good capacity? This tests whether the known noise model alone does the work. If the curve adds nothing over a calibrated residual at a single rate, the trade-off framing is unnecessary. The single-rate ablation of our own model, plus Poisson-scored ARPESNet, settles this directly.
Why not just use DINO feature distances? Ekahana et al. found unsupervised k-means on such features reaches only 50–60% accuracy. That suggests generic image features do not separate subtle electronic states without labels. Distance scores do not localize anomalies in energy and momentum, which physicists need. We still run this baseline, because nearest-neighbour distance is stronger than clustering.
Decisive Experiments
We train on the 46-dataset ARPES library from ARPESNet, holding out entire materials. We evaluate three anomaly families.
The first is injected physical anomalies on held-out spectra: a gap opened at a band crossing, replica bands from zone folding, an extra small pocket, and lifetime broadening that mimics surface degradation. Each is injected at graded strength, which gives exact localization masks and detection-versus-strength curves.
The second is real instrument anomalies. These are the fixed-mode MnBi₂Te₄ and Bi₂Se₃ spectra with detector grids from Liu et al., judged against their swept-mode counterparts as normal.
The third is real physics anomalies. In the Fe₃Sn₂ micro-ARPES maps of Ekahana et al., we train on gapless regions only and score gapped regions, and then repeat with the roles reversed. We also score the graphene flake domain map from ARPESNet against its qualitative domain structure.
Metrics are image-level AUROC and AUPRC, pixel-level AUROC for injected anomalies, false alarms per thousand normal spectra from held-out materials, and scoring throughput per spectrum.
The decisive checks are three. First, normal held-out spectra must reach the floor at tightly clustered rates; if they do not, the score has no meaning. Second, the full-curve score must beat its own single-rate ablation and the capacity ensemble on the Fe₃Sn₂ real-anomaly task, which the model never saw constructed. Third, the anomaly rankings must hold up when we deliberately miscalibrate the detector gain by ±20%, because a method that only works with perfect count calibration would not survive a real beamline.
References
- Iterated learning and communication jointly explain efficient color naming systems
- Cultural evolution via iterated learning and communication explains efficient color naming systems
- Semantic Categories of Artifacts and Animals Reflect Efficient Coding
- Towards Human-Like Emergent Communication via Utility, Informativeness, and Complexity
- Do Autoencoders Need a Bottleneck for Anomaly Detection?
- Video Anomaly Detection with Semantics-Aware Information Bottleneck
- An autoencoder for compressing angle-resolved photoemission spectroscopy data
- Removing grid structure in angle-resolved photoemission spectra via deep learning method
- A machine learning route between band mapping and band structure
- Anomaly-Detection-Driven Screening of Thermodynamic Stability from Composition Descriptors Alone
- Transfer learning application of self-supervised learning in ARPES