Research
Simulation sits between theory and data. We use machine learning to accelerate it in the forward direction, to invert it through simulation-based inference, and to probe what the trained models actually represent.
Generating events from first principles is expensive. We develop neural importance sampling and generative models that learn the structure of matrix elements and phase space, producing fast, statistically reliable samplers that plug into the standard simulation pipeline.
A neural prediction is only useful if its uncertainty is meaningful. We study how to train surrogates and ML-based fits that are calibrated at the level of the physical observables they feed into — a question that is central to machine-learning determinations of parton distribution functions, where the NNPDF approach treats uncertainty estimation as a first-class goal.
Simulation also runs backwards: from data to the underlying physics. We work on inverse problems such as unfolding, and on simulation-based (likelihood-free) inference, where neural networks make parameter estimation possible even when the likelihood cannot be written down explicitly.
Machine-learning models in high-energy physics are powerful but often opaque. We work on making learned representations transparent — connecting what a classifier, a tagger, or a parton-distribution fit encodes to known physical observables, so that an ML result can be understood and trusted, not just used.
A thread through all of it: validity
Across these directions, one question recurs: when can a machine-learning model in physics actually be trusted? We care about methods that are not just accurate but calibrated, robust, and interpretable. This cross-cutting interest connects our work to community efforts on the validation and statistics of ML in fundamental physics — including the PhyStat series and the VERaiPHY initiative on robust AI for physics.
Tools we build on & contribute to
The standard automated tool for LHC event generation. Our methods are built to plug into the MadGraph ecosystem.
The next-generation MadGraph framework, in active development — a target platform for our neural-integration methods.
Neural importance sampling for Monte Carlo integration of matrix elements — a core piece of the group's neural-integration tooling.
Machine-learning determination of parton distribution functions — a key input and methodological touchstone for ML uncertainties and interpretability in QCD.
Networks
The group is part of MCnet, the international Monte Carlo event generator network, through its Milan node.