Hadronization Fingerprints: Speaker-Verification Embeddings for Event-Generator Models

When quarks fly apart in a particle collision, they do not reach the detector as quarks. Below a few GeV the strong force becomes too strong to calculate with standard expansions, and the quarks turn into a spray of ordinary particles called hadrons. This conversion is called hadronization. Physicists simulate it with hand-built models: the string model in Pythia and the cluster model in Herwig. Each has dozens of tunable knobs, and none of them matches precise collider data. We propose treating each hadronization model setting the way speaker recognition treats a person's voice. Each setting is an "identity." Each simulated jet is one short "frame" of evidence. A collection of jets is an "utterance." We train an embedding network that maps a sample of jets to a single vector identifying which hadronization model produced it. The network can then say whether two samples came from the same model, including models it never saw during training.

The motivation is concrete. Speaker recognition has spent a decade learning how to pool many noisy, variable-length frames into a compact identity vector that transfers to unseen speakers. Hadronization physics needs exactly that: a reusable, comparable signature of the non-perturbative regime that can sit next to real data. Today it only has one-off discriminators between pairs of simulations. If the embedding works, real ALEPH data from the Large Electron–Positron collider (LEP) can be placed in the same space as Herwig and Pythia variants. Distances in that space then become a diagnostic of what the generators get wrong.

Core Claim

A sample-level embedding trained with a speaker-verification recipe will separate hadronization models that standard jet observables cannot tell apart, and it will generalize to model settings held out of training. The recipe has three parts. The first is statistics pooling, which keeps the spread of features and not only their average. The second is a two-level hierarchy: hadrons are pooled within a jet, then jets are pooled within a sample. The third is an angular-margin loss over many model settings, which pulls samples of the same setting together and pushes different settings apart.

The specific novelty is the switch from pairwise discrimination to open-set verification over a continuum of generator settings. The closest existing work, Fitting a deep generative hadronization model (HadML v2), fits a generative hadronization model using a Deep Sets discriminator. That discriminator sums per-particle features. It reports an AUC of 0.77 for distinguishing a cluster-smearing (ClSmr) variation in Herwig, compared with traditional observables "which typically hover near 0.5." Its discriminator is trained for one pair of models and then discarded. It produces no coordinate system in which a third model, or the data, can be located.

Why the speech recipe should transfer

The speech side supplies the method. ECAPA-TDNN reaches 0.87% equal error rate on VoxCeleb1. Equal error rate (EER) is the operating point where false accepts equal false rejects. In ECAPA-TDNN's ablations, channel- and context-dependent statistics pooling alone gives a 9.8% relative EER reduction. TitaNet reaches 0.68% EER with cheaper depth-wise convolutions and an angular-margin loss. These results are about speech, and we use them only as evidence that the pooling design matters, not as physics baselines.

The physics premise is that hadronization settings mostly change the fluctuations of soft-particle multiplicities, momentum fractions, and flavor composition. They change the means much less. A sum-pooled set encoder can in principle represent spread, but it has to discover the spread through its nonlinearities. Explicit $[\mu;\sigma]$ pooling hands spread to the network directly.

The two-level structure answers a known speech limitation. Embeddings degrade on short utterances because a variance estimate from few frames is noisy. The equivalent question here is how many jets a sample needs. We expect sample-level verification error to fall roughly as the standard error of the pooled statistics, about $1/\sqrt{n_{\text{jets}}}$, until the embedding saturates. The form comes from averaging $n$ independent jets.

Concretely, a per-hadron encoder $f$ produces features $h_{j,k}$ for hadron $k$ in jet $j$. Attentive statistics pooling gives a jet vector $z_j = [\mu_j;\sigma_j]$. A second attentive statistics pooling over the jets $z_1,\dots,z_n$ gives the sample embedding $e$. Training uses the additive angular margin loss over model-setting labels $y$:

$$\mathcal{L} = -\log \frac{e^{s\cos(\theta_y+m)}}{e^{s\cos(\theta_y+m)}+\sum_{j\neq y}e^{s\cos\theta_j}},$$

Here $\theta_y$ is the angle between $e$ and the learned direction for setting $y$, $s$ is a scale, and $m$ is the margin. Verification of two samples is then a cosine similarity between their embeddings.

Research Program

We would generate $e^+e^-\to$ hadrons at $\sqrt{s}=91.2$ GeV, the Z-pole setup used by HadML v1 and HadML v2. Herwig 7.2.1 supplies a grid of a few hundred cluster-model settings that varies cluster smearing, cluster mass thresholds, and flavor suppression. Pythia supplies a grid of string-model settings that varies string tension and flavor parameters.

Each setting is one training identity. Roughly a fifth of the settings, chosen as contiguous regions of parameter space rather than random points, are held out entirely for open-set testing. Jets are clustered with anti-$k_T$ ($R=0.4$) and SoftDrop grooming ($z_{cut}=0.1$, $\beta=0$), matching the archived ALEPH measurement. That match lets the same pipeline later be applied to the data.

Hadron inputs are four-momenta relative to the jet axis plus a flavor token. The per-hadron encoder is a small transformer that uses the pairwise kinematic attention biases from Particle Transformer, so the encoder itself is not a weak point.

We will analyze three things. First, does the embedding geometry recover the parameter grid smoothly, with nearby settings mapping to nearby vectors, even though training never used parameter values? Second, which hadrons do the attention weights emphasize? We expect soft, wide-angle hadrons, and confirming this would show the embedding is reading non-perturbative physics rather than hard-scatter kinematics. Third, where does ALEPH data land relative to the Herwig and Pythia manifolds?

The third analysis carries a caveat. The ALEPH measurement is corrected to particle level with its own procedure, so the comparison is only as good as the match between our simulated observables and what was unfolded. We will restrict the data comparison to inputs the measurement actually provides.

Baselines

The naive reference is classical jet-substructure observables. These are the observables in the ALEPH paper, such as groomed jet mass, momentum sharing, multiplicity, and energy spectra. They are histogrammed per sample and compared with a chi-square or a gradient-boosted classifier. HadML v2 reports that such observables sit near AUC 0.5 for the ClSmr variation. We will reproduce this on our grid so the claim is tested on the same events.

The strongest existing competitor is the HadML v2 Deep Sets discriminator, which reports AUC 0.77 on ClSmr. We will rerun it as a pairwise classifier on the same simulated events and the same sample sizes, with the same per-hadron feature budget and the same hyperparameter search budget. Beating it means two things. On that same ClSmr pair, our embedding's cosine score must exceed 0.77 AUC. More importantly, on held-out settings the Deep Sets approach has to be retrained per pair, while ours does not. Reaching comparable AUC without retraining would be the main result.

The second strong competitor is a Particle Transformer encoder with its class token, trained with ordinary cross-entropy over the same identities. It reports 86.1% accuracy on JETCLASS versus 84.4% for ParticleNet, for a different task (jet tagging). This baseline isolates whether the speech-specific pieces add anything beyond a strong particle encoder.

Two ablations isolate the key ingredients. The first replaces statistics pooling with mean or sum pooling at both levels, keeping everything else fixed. The second replaces the angular-margin loss with plain softmax. If removing $\sigma$ costs nothing, the core premise that hadronization lives in fluctuations is wrong for this embedding.

Why not just…?

Why not just regress the generator parameters directly, as in simulation-based inference? This is the most serious threat. Within a single generator, a network that predicts string tension or cluster smearing from a sample may tune parameters as well as or better than an embedding. Regression has one structural limit: Herwig's cluster parameters and Pythia's string parameters do not share units or meaning. A regressor therefore cannot place data relative to both families at once, and it cannot flag data that lies off every model's manifold. We will train a regression head on the same encoder anyway. If its parameter estimates on held-out settings are as accurate as a nearest-neighbor readout from our embedding, and the cross-family placement shows nothing beyond what two separate regressors give, then the embedding's value reduces to cross-family comparison. We would report that narrower claim.

Why not just keep training HadML-v2-style pairwise discriminators? They work: 0.77 AUC where observables fail. But each discriminator answers only one question and yields no metric. The held-out-setting experiment settles this directly. If pairwise discriminators retrained per pair clearly beat our zero-shot cosine scores, the embedding is not buying generality at an acceptable cost.

Why not just use better hand-built observables with standard generator tuning? The ALEPH paper shows that all six tested generators (PYTHIA6, PYTHIA8, Sherpa, HERWIG, VINCIA, PYQUEN) leave several-percent discrepancies in soft substructure. HadML v2 shows that standard observables barely respond to cluster-smearing changes. The observable baseline above quantifies that gap on our own grid rather than assuming it.

Proposed Experiments

The decisive check is open-set verification EER on held-out Herwig and Pythia settings, plotted against jets per sample from about 10 to 1000. It is compared with the observable baseline and with retrained Deep Sets discriminators. The secondary check is AUC on the ClSmr pair against the reported 0.77.

The pooling ablation is the key causal test. If the gap between statistics pooling and mean pooling does not grow as samples shrink, the speech analogy fails at its core.

The final, exploratory step embeds ALEPH-matched observables and the simulated samples. It asks whether the data's nearest neighbors are consistent across random seeds, and whether they point to particular parameter directions that standard tunes have missed. If the embedding recovers the held-out parameter grid smoothly but the data sits far from every simulated cluster, that outcome is still informative: it would be a learned, quantitative restatement of the ALEPH paper's conclusion that soft physics modeling needs refinement.