The interface between artificial intelligence (AI) and physics has seen a rapid expansion, with ground breaking advances in both the development and application of machine learning (ML) algorithms designed for domain specific physics problems, and the utilization of tools from physics to better understand theoretical aspects of machine learning. For instance, ML models have enabled signal identification in heterogeneous and 3D detector data, and simulator driven inference and anomaly detection in large and high dimensional datasets, e.g. at the Large Hadron Collider and at the LSST. ML models have also enabled high fidelity generative models as fast approximate simulators or as proposal distributions incorporating physics inductive bias for sampling from complex distributions in lattice QCD, particle scattering amplitudes, and many-body physics. At the same time, theoretical methods from statistical physics and quantum field theory have been applied to understanding neural network learning dynamics and finite network size behaviors.

The first installment of the AI & Physics track at AMLD2020 broadly examined the AI / Physics interface. The AI & Physics track at AMLD2021 will focus on specific subdomains that target improving physics data analysis with the tools of ML and understanding ML models with the tools of physics. The track will contain three sub-sessions covering (i) inference and anomaly detection, (ii) Statistical Physics for understanding ML, (iii) ML for quantum technologies. These topics address key questions of how AI can improve the capability to perform scientific measurements and identify spurious signatures that may be signs of new physical behavior, how ML in the high dimensional limit can be understood through the lens of theoretical physics, and how ML can inform the design and control of quantum experiments.

Machine learning and physics have a symbiotic and transformative influence on each other, leading to profound changes in approaches to physics and ML and thus paving the way to tackle previously intractable problems and to exploit new capabilities. The objective of covering these topics is on the one hand to investigate how ML can optimise the design, control, and information extraction of experiments and to speed up compute-intensive tasks which limit research capabilities. On the other, theoretical tools from physics provide a means to extract information about ML models and their learning dynamics. Thus this track aims to explore the questions: How can we best extract insights from physics data and make use of powerful physics models and high fidelity simulators whilst ensuring such methods are well calibrated? Can we harness the capabilities of ML to automate the delicate design and data collection of quantum experiments? Can we describe the dynamics of neural networks from a first principles understanding?