Speaker
Description
Simulating quantum field theories on the lattice is the most trustworthy first-principles way to study a theory. Computationally, this can be challenging when the lattice becomes large or when the theory approaches a phase transition. In this talk, I will first give a review of recent efforts to apply generative AI to this simulation task. Next, we turn to the question: what have these neural samplers actually learned about the underlying physical theory? Surprisingly, samplers trained with only the bare coupling constants as input can be shown to encode crucial physical information when studied in the right representation. By looking at the trained neural network parameters alone, one can discover phase transitions and even measure critical exponents, suggesting the fascinating interpretation of generative AI models as a new type of physical observable.