SciPy 2025

High-energy physics experiments, such as ATLAS at the Large Hadron Collider (LHC), rely on advanced simulation software to accurately model the dense and complex physical interactions occurring within particle detectors. These simulators generate particle collision events using an implicit likelihood model that is analytically intractable. Traditionally, statistical inference in the presence of intractable models relies on dimensionality reduction techniques, which result in a loss of measurement sensitivity. Simulation-Based Inference (SBI) is an emerging new set of techniques that use deep-learning to construct surrogate models for the analytically intractable likelihood functions directly using high-dimensional data. This leads to significantly improved precision in the measurements of the parameters of the model being tested, making the new set of techniques promising for searches of new physics at the energy frontier.

While powerful, these techniques are also computationally demanding and have thus remained inaccessible for full measurements at the LHC since their first proposal in 2015. However, in 2024, using SBI techniques, we published the most precise measurement of the Higgs boson lifetime using ATLAS data. For this application, several novel ideas were developed that not only scale better to a full LHC measurement, but also offer efficient computational workflows using the Scientific Python ecosystem. The technique and workflow developed are described in four publications, and apart from offering a deeper insight on the Higgs boson properties, have opened up the applicability of SBI techniques for many other precision measurements at the LHC.

In this talk, I will describe the significant role that the open-source Scientific Python ecosystem played in the development of the full framework at the ATLAS experiment. Using a live demo, I will demonstrate how we went from ideas to a working at-scale implementation, using tools like TensorFlow and JAX. I will highlight the various challenges, both foreseen and un-foreseen, that we encountered along the way and how the versatility and power of the Python tools helped in overcoming each of them and making the final measurement of the Higgs boson lifetime using the full ATLAS model possible.

This talk will be of interest to people doing precision measurements in a complex likelihood-free setting, which could include any domain from particle physics to biology. Attendees will primarily learn about the versatility of libraries like JAX for efficient statistical inference, including the use of just-in-time compilation, auto-vectorization, and auto-differentiation. For people who don’t fall into this target audience, they will learn about how the various available tools in the Scientific Python ecosystem are being used in some of the cutting-edge research and what future developments can help tackle emerging challenges.