AllShowers unifies calorimeter shower simulation for electrons, photons, and charged and neutral hadrons in one generative model, surpassing prior single-particle models on hadronic showers

The work introduces AllShowers, a unified generative model that simulates calorimeter showers across multiple particle types with a single model rather than separate networks trained per particle species. Traditional surrogate models for calorimeter shower modeling train separate networks for each particle species, limiting scalability and reuse; AllShowers consolidates multi-particle shower generation into one model. The abstract states the model is trained on a diverse dataset of simulated showers in the highly granular ILD detector and generates showers for electrons, photons, and charged and neutral hadrons across a wide range of incident energies and angles without retraining.

AllShowers uses a continuous normalizing flow model with a Transformer architecture to generate complex spatial and energy correlations in variable-length point cloud representations of showers. Combining continuous normalizing flows with a Transformer for point-cloud shower representations lets the model handle variable-length point clouds and capture their spatial and energy correlations. The abstract specifies the model type (continuous normalizing flow), architecture (Transformer), and data representation (variable-length point clouds), but gives no network size, training sample count, or computational cost figures.

The work introduces three key design elements: a layer embedding, a custom attention masking scheme, and a shower- and layer-wise optimal transport mapping. The layer embedding lets the model learn all relevant calorimeter layer properties; the custom attention masking reduces computational demands while introducing a helpful inductive bias; the shower- and layer-wise optimal transport mapping improves training convergence and sample quality. The abstract lists these three designs under 'Key innovations' and describes their roles, but provides no ablation studies or quantitative comparison details.

Beyond unifying multi-particle generation, AllShowers surpasses the fidelity of previous single-particle-type models for hadronic showers. Prior single-particle-type models are surpassed on hadronic showers, indicating the unified model does not sacrifice hadronic shower quality. The abstract states this fidelity improvement directly, but gives no specific metric, numerical value, or statistical uncertainty.

The work targets fast simulation of calorimeter showers in collider experiments, applies to the setting of the highly granular ILD detector, covers electrons, photons, and charged and neutral hadrons, and works across a wide range of incident energies and angles without retraining. It makes it possible for a single model to replace multiple surrogate models trained per particle species, offering direct value to readers working on detector simulation, event reconstruction, and machine learning surrogate models.

The abstract gives no specific evaluation metrics, sample sizes, training costs, or numerical comparisons with baseline models, so the magnitude of the claimed improvement over previous single-particle-type models cannot be judged from the abstract. The individual contributions of the layer embedding, attention masking, and optimal transport mapping also need ablation studies in the main text to confirm. In addition, results are based on simulated data in the ILD detector, and applicability to other detector geometries or real data remains an open question.