Davide Valsecchi

Postdoctoral researcher at ETH Zurich and CERN

Postdoctoral Researcher at ETH Zurich, based at CERN on the CMS experiment at the Large Hadron Collider. I work on three connected fronts: generative machine learning research, precision calibration of a large particle detector, and high-performance ML inference in production.

I specialize in probabilistic modeling, simulation-based inference and generative deep learning, with a focus on normalizing flows and transformers. My research designs surrogate models for expensive simulations and architectures for unbinned likelihood estimation, and brings transformer and graph-based models into detector reconstruction.

Since 2025 I co-coordinate the CMS Electron/Photon Physics Object Group, about 30 people delivering calibrated electrons and photons to every analysis in the collaboration. I set the reconstruction and calibration strategy, and I use the role to modernize both: simulation-based inference with normalizing flows replaces thousands of binned fits with continuous, multivariate calibrations, and transformer and graph-based models replace the legacy clustering algorithms.

As Machine Learning Software Coordinator of CMS (2022–2025) I led the integration of PyTorch and ONNXRuntime inference into the multi-million-line C++ reconstruction software, and designed a zero-copy interface for ML inference on GPUs (ACAT 2025, FastML 2025).

I created and lead PocketCoffea, an open-source framework for reproducible, large-scale analysis of collider data (Python, Awkward Array, Dask, HTCondor), used by >10 large CMS analysis efforts and about 10 active core contributors to process the CMS data and simulation campaigns on HPC clusters and the LHC computing grid. I prototype fast, with agent-based development and LLM coding agents as daily tools.

On the physics side, I lead the CMS search for the Higgs boson produced with a top-quark pair and decaying to b-quarks (ttH, H→bb) with LHC Run 3 data, aiming at the first observation of this process in this final state. My PhD work gave the first evidence of vector boson scattering in the semileptonic channel at the LHC.

Interests: generative models (normalizing flows, flow matching), simulation-based inference and uncertainty quantification, surrogate models for simulation, graph neural networks and transformers for sparse sensor data, ML in production on HPC and GPU clusters, LLM agents for scientific software development, open-source scientific software.

Links: CV (PDF) · GitHub · Inspire HEP · ORCID · LinkedIn · dvalsecchi@ethz.ch

Highlights

  • Factorizable Normalizing Flows — a generative model whose density deforms as a function of continuous parameters, with cost linear in the number of parameters. Turns the classic “template morphing” of physics fits into a differentiable, high-dimensional tool (arXiv:2606.30489, stat.ML, 2026).
  • Uncertainty-aware simulation-based inference — fully profiled unbinned fits that measure a whole distribution instead of scalar parameters, with statistical and systematic uncertainties in one likelihood (arXiv:2602.13184, 2026).
  • Simulation-to-data correction with one flow — a single conditional normalizing flow that morphs simulated distributions onto real data (Comput. Softw. Big Sci. 8, 15, 2024).
  • Generative surrogate models for the Matrix Element Method — transformers and flows sample a high-dimensional integration space, making a statistically optimal but computationally prohibitive method practical (MEMFlow, EuCAIF 2025).
  • PocketCoffea — creator and lead developer of an open-source columnar analysis framework that scales from a laptop to HTCondor and Dask clusters (GitHub · docs).
  • ML inference on heterogeneous hardware — zero-copy Structure-of-Arrays to PyTorch tensors, thread- and stream-safe scheduling, ahead-of-time compiled models in production C++ workflows (ACAT 2025, FastML 2025).
  • Graph neural networks for sensor data — first GNN-based energy clustering in the CMS calorimeter, more robust to noise and pile-up than the legacy geometric algorithm (J. Phys. Conf. Ser. 2438 012077, 2022).
  • Evidence for vector boson scattering — PhD thesis result: a deep-learning signal extraction gave the first evidence of this rare process in its semileptonic channel at the LHC (Phys. Lett. B 834, 137438).