GAN-based data augmentation for rare and exotic hadron searches in Pb--Pb collisions in ALICE

This work presents a feasibility study aimed at enhancing the reconstruction sensitivity for rare heavy-flavour hadrons in Pb-Pb collisions in the ALICE experiment, using the $Ξ_{c}^{+}$ baryon as a benchmark. The $Ξ_{c}^{+}$ baryon has a low rate of production and some complex decay topologies as for instance the decay $Ξ_{c}^{+} \rightarrow Ξ^{-} + π^{+} + π^{+}$ considered in this work. Traditional simulation workflows involving event embedding and full detector response are computationally expensive and statistically limited, especially for rare signals. This study represents the first exploration of generative models within the heavy-flavour programme of ALICE. It uses a dataset of reconstructed physics quantities, such as momenta, positions, and decay vertex coordinates of $Ξ_{c}^{+}$ decay products in Pb-Pb collisions as input features, derived from augmented ALICE Monte Carlo simulations. Such features will serve as a training set for Generative Adversarial Networks (GANs) designed to generate statistically significant synthetic signal samples without the need for additional full simulations. While $Ξ_{c}^{+}$ serves as a benchmark, the broader objective is to enable searches for exotic heavy-flavour hadrons or other exotic states with complex decay patterns. By leveraging GAN-based augmentation, this approach supports rare-signal extraction in computationally demanding analyses and opens the way to broader applications of generative models in the ALICE heavy-flavour programme.

💡 Research Summary

The paper presents a feasibility study that explores the use of Generative Adversarial Networks (GANs) as a data‑augmentation tool for rare heavy‑flavour hadron searches in the ALICE experiment, with the Ξc⁺ baryon serving as a benchmark case. Heavy‑flavour and exotic hadron measurements in ultra‑relativistic Pb–Pb collisions are limited by extremely low production rates and overwhelming combinatorial background. Traditional Monte‑Carlo (MC) workflows rely on event embedding and full detector simulation, which are computationally intensive and statistically insufficient for very rare signals.

The authors focus on the decay channel Ξc⁺ → Ξ⁻ + π⁺ + π⁺, a cascade topology that creates multiple secondary vertices and thus poses a substantial reconstruction challenge in the high‑multiplicity environment of Pb–Pb collisions. They extract a set of reconstructed observables from MC‑generated signal candidates – decay lengths, pointing angles, distances of closest approach (DCA) to the primary vertex, and the kinematic variables (pT, η, φ) of the three decay products – and treat these as a high‑dimensional feature vector.

A standard GAN architecture is employed: a generator maps random Gaussian noise to synthetic feature vectors, while a discriminator learns to distinguish generated vectors from real MC vectors. Training proceeds for several thousand epochs. Early in training the generated distributions deviate strongly from the MC reference, as expected. Convergence is monitored using both loss curves and a statistical validation based on the Kolmogorov‑Smirnov (KS) test. For each observable the KS p‑value is computed between the MC reference and the generated sample; p‑values above 0.05 are taken as evidence of statistical compatibility. After roughly 5 × 10³ epochs, most observables achieve p‑values well above this threshold, indicating that the GAN reproduces the marginal distributions.

Beyond one‑dimensional agreement, the authors examine two‑dimensional correlations (e.g., DCA vs. pointing angle, decay length vs. pT) and find that the synthetic data capture the shape and density of the MC distributions, with only a few isolated outliers. The stability of the adversarial training is demonstrated by the smooth evolution of generator and discriminator losses, showing no signs of mode collapse over the entire training period.

The study demonstrates that GAN‑generated samples can be used to augment the limited pool of rare signal events without the need for additional full detector simulations. Potential applications include (i) expanding the training set for machine‑learning classifiers (BDTs, deep neural networks) that separate signal from background, (ii) testing analysis strategies under realistic heavy‑ion conditions with a larger synthetic signal sample, and (iii) reducing the overall CPU time required for heavy‑flavour analyses.

The authors acknowledge several limitations. The current approach uses only reconstructed observables, not raw tracking information, so any detector‑level effects not captured in the MC may be propagated unchanged. Moreover, the GAN inherits any systematic biases present in the underlying MC model. The study is limited to a single decay channel; extending the method to other exotic states with even more complex topologies will require additional validation.

Future work is outlined: incorporating conditional GANs or normalizing‑flow models to allow control over specific kinematic regions, expanding the feature set to include low‑level detector quantities, and performing cross‑checks with real data to quantify systematic uncertainties. The authors also plan to apply the GAN‑augmented samples in a full physics analysis, evaluating improvements in signal significance, background rejection, and robustness against analysis variations.

In conclusion, the paper provides the first concrete demonstration that GAN‑based data augmentation can faithfully reproduce the multi‑dimensional distribution of reconstructed heavy‑flavour signal candidates in ALICE. This technique offers a promising route to alleviate the computational bottleneck of full MC simulations, thereby enhancing the sensitivity of searches for rare and exotic hadrons in heavy‑ion collisions.