Speaker
Description
High-precision three-dimensional magnetic field mapping is crucial for precision muon experiments, particularly in beam transport, injection, and storage regions where direct sensor access is often restricted. We present a Physics-Informed Neural Network (PINN) framework that incorporates Maxwell’s equations (∇·B = 0 and ∇×B = 0) directly into the training loss. The method features a dual-constraint strategy enforcing physics consistency at both collocation points and measurement locations, along with a three-stage optimization (AdamW, LBFGS, and Residual-based Adaptive Refinement).
Biot-Savart simulations demonstrate reconstruction accuracy reaching ~10^{-4}, representing a tenfold improvement over prior PINN methods, with strong robustness against sparse data and noise. Laboratory experiments using a custom coil assembly under ambient conditions achieve sub-percent relative accuracy (down to 10^{-3} level).
In collaboration with J-PARC researchers, a small-scale application to a 3D-injection simulation dataset achieved approximately 1% prediction error. These results highlight the potential of PINNs as a powerful tool for high-precision magnetic field reconstruction in muon g−2/EDM and related experiments.
