PhD candidate · University of Hamburg
Open to postdoc positions starting in 2027

I work on generative models for the simulation of particle showers in calorimeters. Detailed simulation with Geant4 dominates the computing budget of collider experiments, and learned surrogates have so far been tied to the one detector whose data trained them.

My thesis makes one generative shower model reusable across calorimeter geometries. Pre-trained on several detectors, it adapts to a new one with two orders of magnitude fewer Geant4 showers than a model trained from scratch.

Lorenzo Valente
CV INSPIRE-HEP Scholar ORCID GitHub Hugging Face LinkedIn email

About

I am a PhD candidate at the Institute of Experimental Physics of the University of Hamburg, in the group of Gregor Kasieczka. Before Hamburg I studied physics in Bologna, where my master's thesis was on real-time anomaly detection at CMS. I am also a classical pianist.

Research

Multi-geometry pre-training

I extend a flow matching transformer for shower point clouds with conditioning on the calorimeter geometry and pre-train it across several detectors. Fine-tuned on a detector it has never seen, it needs two orders of magnitude fewer Geant4 showers to match a model trained from scratch on a full dataset (arXiv:2608.18233).

Sliced Wasserstein distance to Geant4 against the number of training showers on FCCee-ALLEGRO, for the longitudinal profile, the radial profile, the cell energy spectrum and their geometric mean, comparing training from scratch with fine-tuning from the LEMURS, SimpleBox and Mini pre-trainings
Sample efficiency on FCCee-ALLEGRO, a calorimeter outside the pre-training pool. Sliced Wasserstein distance to Geant4 as a function of the number of training showers, for training from scratch and for fine-tuning from each pre-trained model. Bands span the mean ± one standard deviation over five seeds, and the dashed line is the Geant4-versus-Geant4 statistical resolution. From arXiv:2608.18233.

Cross-geometry transfer

The first step was to show that a shower model trained on one calorimeter transfers to another with a fraction of the data (JINST 21 (2026) P07037).

Datasets

The pre-training data are produced with Geant4 and DD4hep for a family of sampling calorimeters, from simulation to training-ready HDF5 (multi-calorimeter-dataset). The datasets themselves are on the University of Hamburg research data repository. Along the way I develop extensions to ddFastSim, the DD4hep fast simulation framework.

Now

A numerical geometry descriptor as conditioning, towards a model that generalises to a new detector without any fine-tuning.

Publications

Journal articles and preprints

Conference proceedings

Full list on INSPIRE.

Talks

Lorenzo Valente giving the invited talk at LHCP 2026
Presenting at LHCP 2026, Sorbonne University, Paris, May 2026.

Conference talks

Other talks

Teaching

Code and data

SimpleBox and LEMURS datasetsPoint cloud shower datasets for multi-geometry pre-training, DOI 10.25592/uhhfdm.19103. multi-calorimeter-datasetGeant4/DD4hep production of the SimpleBox and LEMURS shower datasets. AllShowersFlow matching transformer for shower point clouds, multi-geometry branch. PointCountFMPer-layer point count model, multi-geometry branch. Weights on Hugging FacePre-trained models of the multi-geometry paper. CaloTransferCode of the JINST cross-geometry transfer paper. JointVAE4ADJoint continuous and discrete VAE for anomaly detection. Autoencoder-for-FPGAQuantization-aware autoencoder deployed with hls4ml.