Learning to Read Resonances Across Reactions with Phase-Randomized Meta-Training

Many short-lived particles studied in nuclear physics, called hadrons, have nearly the same mass and very broad, overlapping signals. When several of them decay to the same final particles, their quantum amplitudes add before being squared. The measured spectrum can therefore show shifted peaks, flat plateaus, or dips instead of clean bumps. Neural networks trained on simulated data have begun to read particle properties directly from such spectra. The first such work, a network that identifies $\Sigma$ baryon resonances in $K^-p$ scattering, loses accuracy quickly as overlapping states are added: 98.5% with one extra resonance, 94.8% with two, and 82.5% with three. It is also trained and tested on a single reaction from a single theoretical simulator (Σ Resonances from a Neural Network-based Partial Wave Analysis). We argue that this network fails in the way meta-learning researchers call memorization: it learns one fixed map from spectrum shape to particle properties, valid only for the one simulator it saw.

The proposal rests on a physics fact. A resonance's pole position (its mass and width) is universal, the same in every reaction where the particle appears. Its production strength and relative phase depend on the reaction. We define each reaction-and-simulator configuration as a separate "task" and randomize the reaction-dependent phases and couplings coherently, at the amplitude level, across tasks. A given spectrum then becomes ambiguous unless the network looks at a small calibration sample from the same reaction. This is exactly the condition that Rajendran et al. showed forces a meta-learner to adapt instead of memorize. Unlike their label shuffling, it never assigns a physically wrong label.

Core Claim

Gradient-based meta-learning, such as MAML, trains a model whose starting weights can be tuned to a new task in a few gradient steps on a small labeled "support set" and then scored on a held-out "query set." Meta-Learning Requires Meta-Augmentation showed that this fails when one fixed model can solve all training tasks. The cure is an augmentation that makes the label harder to predict from the input alone while leaving it predictable once the support set is seen. In vision, this is done by shuffling class labels. Shuffling particle quantum numbers or masses would violate conservation laws, and the research file flags this as an open tension.

Our claim is that amplitude-level randomization of process-dependent nuisances is the physically legitimate version of this augmentation for resonance extraction. A task $\tau$ generates events $x$ from

$$\mathcal{A}_\tau(x) = \sum_r g_{\tau r}\, e^{i\phi_{\tau r}}\, F_r(s; M_r,\Gamma_r)\, T_r(\Omega) + B_\tau(x),$$

where $M_r,\Gamma_r$ are the shared pole parameters that form the labels. The coupling magnitude $g_{\tau r}$ and phase $\phi_{\tau r}$ are drawn per task, as is the background amplitude $B_\tau$, which interferes with the resonances. $F_r$ is the lineshape and $T_r$ the angular factor, both from the standard isobar decomposition used in Dalitz plot analysis of D_s^+ → K^+K^-π^+. The measured density is $|\mathcal{A}_\tau|^2$, so the cross terms make the same poles look very different from task to task.

The design target, in the language of conditional entropy (how uncertain the label is given what the model sees), is

$$H(y \mid x) \gg H(y \mid x, S_\tau),$$

where $y$ is the set of pole parameters and quantum numbers and $S_\tau$ is the task's support set. In words, the event sample alone should be ambiguous, while the sample plus a small calibration set from the same process should not be.

We predict two things. First, a meta-learner trained this way will extract poles from held-out reactions and held-out lineshape models with lower bias and better-calibrated uncertainties than a single network trained on pooled data. Second, the gap will widen as the number of overlapping resonances grows.

Research Program

We will build a generator suite of three-body and quasi-two-body decays. Each generator is seeded with the resonance content of a published analysis in the research file. These are the six-component $D_s^+ \to K^+K^-\pi^+$ decomposition, whose fit fractions sum to about 130% because of interference (Mitchell et al.); the interfering $f_0(1710)$ in $J/\psi \to \omega K^+K^-$ (Ablikim et al.); the broad $0^{-+}$ and $2^{++}$ states in $J/\psi \to \gamma K^+K^-\pi^+\pi^-$ (Bai et al.); the $X(1840)$ region in $J/\psi \to \gamma 3(\pi^+\pi^-)$ (Observation of a structure at 1.84 GeV); and an isobar re-implementation of the $K^-p \to \pi^0\Lambda$ setting. We will write our own generators, since the Effective Lagrangian code used by Shi et al. is not available to us.

A task fixes three things: a reaction, a lineshape convention (a simple Breit–Wigner peak shape, or a variant with different form factors and barrier factors), and a background model. Within a task, couplings, phases, and pole parameters are sampled. The support set is a small labeled simulated sample from that task, standing in for the channel-specific simulation that an experiment always has for its own detector. The query set is a fresh event sample to be analyzed.

The network outputs the quantum-number content as a classification and the masses and widths as regression targets. It also produces a per-parameter uncertainty, so that calibration can be scored.

To control cost, adaptation is restricted to the final layers, in the style of ANIL and first-order MAML, so we avoid differentiating through the full adaptation trajectory. This choice follows the DropGrad finding that later layers carry the task-specific, memorizable features (Regularizing Meta-Learning via Gradient Dropout).

On top of the phase-randomized tasks, we will compare three existing meta-regularizers, each used as published and treated as a secondary factor. The first is gradient dropout, which randomly masks the gradients used during adaptation (Tseng et al.). The second is inverted regularization inside the adaptation step, which adds a negative penalty there and a normal one at the outer level (Improving Generalization of Meta-Learning with Inverted Regularization at Inner-Level). The third is Meta Dropout, a learned noise generator that perturbs hidden features (Meta Dropout).

The inverted-regularization comparison also tests a tension the file identifies. Physicists routinely add positive constraints to stabilize individual fits, while Wang et al. show that positive regularization inside the adaptation step hurts generalization across tasks. We will measure which direction helps here.

Baselines

The naive reference point is the Shi et al. recipe. It is a single convolutional network trained by ordinary supervised learning on one generator, with static input noise of the kind they call Artificially Expanded Data (AED). We will apply it to every task family and evaluate it on held-out reactions without any adaptation.

The strongest reported number for this question is theirs: 98.5%, 94.8%, and 82.5% test identification accuracy for one, two, and three additional resonances (Shi et al.). That number comes from same-generator testing on their own simulator. We will reproduce the protocol on our isobar version of $K^-p$ and treat our reproduced figure as the comparable reference. Beating the baseline means two things. On held-out reactions and held-out lineshape conventions, our method must give higher three-resonance identification accuracy and smaller pole-mass bias. Its uncertainties must also be calibrated: when we compute each estimate's error divided by its own reported uncertainty (the "pull"), the spread of these pulls should be between 0.9 and 1.1. The method must not lose accuracy on the in-distribution $K^-p$ setting.

The strongest machine-learning competitor is a pooled, domain-randomized supervised network. It is the same architecture, trained by ordinary supervised learning on exactly the same phase-randomized task mixture, with the same number of simulated events and the same hyperparameter search budget, and it has no support set. This is the comparison that isolates whether adaptation matters at all.

A second strong competitor is a context-conditioned network. It receives an embedding of the support set as an extra input and adapts by conditioning rather than by gradient steps, with matched parameter count and training budget.

The classical reference is an unbinned maximum-likelihood isobar fit on each query sample. It is given the true model family and 50 random restarts to cope with the many local minima the file documents. We are not aware of a reported number on these synthetic configurations; no reported result was found on Lacuna for multi-start likelihood fits on these synthetic configurations, so we will measure it ourselves.

The key ablation replaces amplitude-level randomization with probability-level randomization. In that variant, resonance contributions are added incoherently and reweighted, in the manner of MetaMix interpolation (Improving Generalization in Meta-learning via Task Augmentation; MetaMix). If this incoherent variant performs as well as the coherent one, our central claim about interference is wrong.

For context, these meta-regularizers have known effects on standard benchmarks. Meta-augmentation reduced Pascal3D pose-regression error from 2.41 to 1.61 MSE (Rajendran et al.). Gaussian DropGrad raised mini-ImageNet to CUB 5-shot accuracy from 45.56% to 51.05% (Tseng et al.). Neither result has been tested on physics amplitudes.

Why not just…?

Why not just train one big supervised network on the randomized simulations? This is the real threat. With enough events, interference patterns across a Dalitz plot (the two-dimensional map of pair masses in a three-body decay) can identify the phases from the query data alone. In that case pooled training would learn the right answer without any support set. We expect meta-learning to win only where the channel-specific nuisances are not identifiable from the query sample, namely background shape, lineshape convention, and acceptance. We will settle this by sweeping the query sample size from hundreds to tens of thousands of events. If the pooled network matches the meta-learner at the event counts typical of real analyses, from 1,516 events in Bai et al. to 14,400 in Mitchell et al., we will report that and recommend domain randomization instead.

Why not just feed the support set as an input rather than doing gradient adaptation? A context-conditioned network is cheaper and avoids second-order gradients, and it may well be enough. It is included as a matched baseline. If it ties, the contribution reduces to the phase-randomized task design, which is the part we consider essential. The gradient-based form would then be dropped.

Why not just run a careful multi-start likelihood fit? Likelihood fits remain the standard and resolved the six-component $D_s^+$ decomposition and the $X(1840)$ at $7.6\sigma$. However, the file documents flat directions, convergence failures, and pole-mass shifts of about ±50 MeV for broad states when the lineshape model is wrong (Bai et al.). We will compare the two approaches under deliberate model mismatch: data generated with one lineshape convention, analyzed assuming another. We expect the fit to show biased pulls there and the adapted network to show smaller bias. If the fit's bias is not larger, the network offers only speed.

Why not just add dropout or weight decay? Tseng et al. and Wang et al. report that these single-task regularizers disrupt or weaken adaptation across tasks, which is why they are outside the main comparison. We still include a weight-decay-plus-dropout arm of the pooled baseline so the result is checked on our data.

Proposed Experiments

The first decisive check is a memorization diagnostic adapted from Rajendran et al. We swap the support set for that of a different task. A model that truly adapts should lose accuracy sharply when given the wrong support; a model that memorizes should be unaffected. This tells us directly whether phase randomization changed how the network behaves.

The second check is a generalization grid. Training and test differ in reaction, lineshape convention, and number of overlapping resonances, from one to three. We score identification accuracy, pole-mass bias, and the spread of pulls in every cell. The main prediction is a smaller drop than 98.5% to 82.5% as overlaps increase, measured on held-out reactions rather than on the training simulator.

The third check is the coherent-versus-incoherent augmentation ablation, together with a regularizer comparison of gradient dropout, inverted adaptation-step regularization, and Meta Dropout. The regularizer comparison asks whether the sign of the penalty inside the adaptation step matters for physics tasks as it does for images.

Finally, as a consistency test, the network is adapted to a $D_s^+ \to K^+K^-\pi^+$ configuration generated with the published CLEO-c fit fractions. It should recover those fractions within their quoted uncertainties, including the small $f_0(1370)$ component (4.3%) that interferes with neighboring states. Analysis of real experimental data is out of scope unless open data become available.

References

  1. $Σ$ Resonances from a Neural Network-based Partial Wave Analysis on $K^-p$ Scattering
  2. Meta-Learning Requires Meta-Augmentation
  3. Dalitz plot analysis of<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msubsup><mml:mi>D</mml:mi><mml:mi>s</mml:mi><mml:mo>+</mml:mo></mml:msubsup><mml:mo>→</mml:mo><mml:msup><mml:mi>K</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>K</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:msup><mml:mi>π</mml:mi><mml:mo>+</mml:mo></mml:msup></mml:math>
  4. Study of <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" altimg="si1.gif" overflow="scroll"><mml:mi>J</mml:mi><mml:mo stretchy="false">/</mml:mo><mml:mi>ψ</mml:mi><mml:mo>→</mml:mo><mml:mi>ω</mml:mi><mml:msup><mml:mi>K</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>K</mml:mi><mml:mo>−</mml:mo></mml:msup></mml:math>
  5. Partial wave analysis of J/ψ→γ(K+K−π+π−)
  6. Observation of a structure at<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mn>1.84</mml:mn><mml:mtext> </mml:mtext><mml:mtext> </mml:mtext><mml:mi>GeV</mml:mi><mml:mo>/</mml:mo><mml:msup><mml:mi>c</mml:mi><mml:mn>2</mml:mn></mml:msup></mml:math>in the<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mn>3</mml:mn><mml:mo stretchy="false">(</mml:mo><mml:msup><mml:mi>π</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>π</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:mo stretchy="false">)</mml:mo></mml:math>mass spectrum in<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>J</mml:mi><mml:mo>/</mml:mo><mml:mi>ψ</mml:mi><mml:mo>→</mml:mo><mml:mi>γ</mml:mi><mml:mn>3</mml:mn><mml:mo stretchy="false">(</mml:mo><mml:msup><mml:mi>π</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>π</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:mo stretchy="false">)</mml:mo></mml:math>decays
  7. Regularizing Meta-learning via Gradient Dropout
  8. Improving Generalization of Meta-Learning with Inverted Regularization at Inner-Level
  9. Meta Dropout: Learning to Perturb Latent Features for Generalization
  10. Improving Generalization in Meta-learning via Task Augmentation
  11. MetaMix: Improved Meta-Learning with Interpolation-based Consistency Regularization