TL;DR
CaloCHILL (Calorimeter Hits In Learned Latents) is a novel generative model that accelerates hadronic shower simulation by generating a compact latent summary of a shower before expanding it into a full-resolution point cloud. It achieves a 600x speedup over its predecessor while being the first model to successfully simulate the complex 5D timing structure of hadronic showers in high-granularity calorimeters.
What Problem Does This Paper Address?
In High Energy Physics (HEP), simulating how particles interact with detectors is a massive computational bottleneck. The "gold standard" software, Geant4, is extremely accurate but painfully slow, especially for calorimeters where a single particle can trigger thousands of individual energy deposits (hits).
As we design future experiments like the FCC-ee, detector granularity is increasing, making traditional simulation even more expensive. While "fast simulation" surrogates exist, many struggle to scale because they generate hits one by one. Furthermore, most existing fast simulators ignore the timing of hits—a crucial feature for future detectors to disentangle overlapping particle interactions in high-luminosity environments.
Methods
The authors introduce a three-stage factorized generation pipeline that avoids the "point-by-point" bottleneck:
- LayerFM: A flow-matching model that predicts "macro" features: the number of hits and the earliest hit time for each of the 78 calorimeter layers. This sets the overall longitudinal skeleton of the shower.
- Latent Mean Flow (LMF): This stage generates a sequence of latent vectors from Gaussian noise. Instead of thousands of points, it only needs to produce a few hundred latent "summaries."
- VAE Decompressor: A Variational Autoencoder (VAE) takes the latent sequence and expands it into the final point cloud, assigning each hit a position ($x, z$), energy ($E$), and time ($t$).
To keep the model efficient, the authors utilize Sparse Attention. Since hadronic showers develop along a central axis, hits are mostly correlated with their immediate neighbors. By using local attention masks (Figure 3), the model avoids the $O(N^2)$ computational explosion of standard transformers.
Datasets
The model was trained and evaluated on the CaloHadronic dataset, consisting of:
- Particles: Single positive pions ($\pi^+$).
- Energy Range: 10 GeV to 90 GeV.
- Detector: A simulation of the International Large Detector (ILD), featuring a 30-layer electromagnetic calorimeter (ECal) and a 48-layer hadronic calorimeter (HCal).
- Features: Each hit includes 5D information: $(x, y, z, E, t)$.
Key Insights and Ideas
- Generation in Latent Space: By compressing the shower into a sequence of latent vectors, the generative stage (LMF) becomes independent of the total number of hits. The heavy lifting of "expanding" the shower is done only once by the VAE decompressor.
- Macro-Micro Separation: Predicting the number of hits per layer before generating individual points (via LayerFM) significantly improves the fidelity of the longitudinal profile.
- The "Slow" Component: Hadronic showers aren't instantaneous; they have a prompt electromagnetic core followed by a "slow" tail from neutron scattering. CaloCHILL is the first to capture this temporal evolution across both ECal and HCal sections.
- Physics-Informed Loss: The VAE is trained using a Sliced Wasserstein Distance layer-by-layer. This allows the model to compare unordered sets of points (distributions) much more efficiently than standard Chamfer or Earth-Mover distances.
Key Results
- Speed: CaloCHILL is 600x faster than its predecessor, CaloHadronic. When compared to Geant4 at equal hardware cost, it provides a ~50x speedup.
- Accuracy: It outperforms previous state-of-the-art models like AllShowers in most metrics, particularly in reproducing the longitudinal energy profile (Figure 6).
- Timing Fidelity: The model accurately reproduces the arrival time distribution, including the prompt front and the slow neutron component extending out to 200 ns.
Qualitatively, the generated showers are nearly indistinguishable from Geant4 "truth" simulations, as seen in the event displays:
Limitations and Future Work
- Fixed Geometry: The current model is trained on a "particle gun" firing vertically. Future versions will need to handle varying incident angles and different particle species.
- Compression Artifacts: The authors noted a slight deficit in energy very far from the shower axis and an excess of low-energy cells, which are inherent to the VAE's compression.
- Dynamic Latents: Currently, the number of latent vectors ($d_s$) is fixed. Scaling $d_s$ dynamically with the number of hits in a shower could further improve accuracy for very high-energy events.
Summary updated