Speaker
Description
Accurate prediction of sound energy decay in enclosed spaces is essential for acoustic design in performance halls, classrooms, and recording studios. Statistical room-acoustic theory, such as the Sabine reverberation formula, provides fast but spatially limited estimates, while wave-theoretical simulations are physically accurate but computationally expensive. This work presents a Physics-Informed Neural Network (PINN) trained to solve the acoustic diffusion equation, using finite-difference time-domain (FDTD) simulations for supervised learning. The network takes spatial position and time as inputs and predicts sound energy density without requiring re-simulation. The maximum likelihood enforces the governing partial differential equation, boundary conditions, and initial conditions alongside supervised data. Preliminary results show close agreement with FDTD ground truth across the full decay range, accurately reproducing the temporal decay process at the receiver positions. Full temporal coverage of the decay process provides stronger training than spatially dense but temporally sparse sampling. The current work also explores Bayesian probabilistic evaluations to quantify uncertainties during training with the goal of generalizing across room geometries and absorption coefficients without retraining.