New Domain Adaptation Method Improves Physics Simulations

Domain adaptation is a common technique used to bridge the gap between neural networks trained on simulated data and their application to real-world experimental data. The underlying assumption is often that the two domains differ only in "nuisances" (e.g., detector response) and that the quantity of interest (e.g., an energy spectrum) is distributed identically. However, in physics, these assumptions frequently break down: simulations can be physically inaccurate, and the target quantity's distribution is often the very measurement being sought. This research investigates the consequences of such mismatches using a toy air-shower benchmark, where detector nuisances, physical simulation shifts, and energy-spectrum shifts can be independently controlled. It demonstrates that standard adversarial adaptation methods, while handling conditional shifts, can introduce uncontrolled biases by aligning differing spectra based on simulation priors. To overcome this, the paper proposes "adaptive domain adaptation." This new method reweights simulated events to specifically focus the domain adaptation process on genuine physical mismatches, rather than inadvertently aligning distributions that are fundamentally different. It also provides a label-free model selection rule to identify the optimal operating point, ensuring that the predicted spectrum is robust against model training configurations and simulation inaccuracies.