PAWS: A Physical Acoustic Wave Simulation Dataset for Sound Modeling and Rendering

We present a Physical Acoustic Wave Simulation (PAWS) dataset for room acoustic simulation based on wave acoustics simulation. Previous geometrical acoustics-based room acoustic simulation methods have difficulty in modeling diffraction and low-frequency sounds. Many previous acoustic datasets only provide impulse responses at limited positions without offering the entire acoustic pressure field. There are limited acoustic field datasets available, and they mainly focus on single or a limited number of scenes. In this work, we simulate a large-scale wave acoustics-based room acoustic dataset, PAWS, using acoustic finite-difference time-domain (FDTD) and k-space solvers. The PAWS dataset contains acoustic pressure p(x,y,z,t) data at 65,536 grid samples in space. PAWS consists of 200 scenes randomly sampled from shoebox rooms with subrooms and obstacles, polygon rooms, and real-world scenes from the Replica dataset. We randomly place source and obstacles in each scene and run the simulation. In total, PAWS consists 2,000 data samples. The PAWS dataset is designed for benchmarking neural network-based methods for fast acoustic fields prediction in new scenes. We propose mean squared error (MSE) and signal-to-noise ratio (SNR) to evaluate the acoustic pressure estimation systems. We provide a physics-informed neural network (PINN)-based benchmark system to predict the acoustic pressure in the sound field. We release both the PAWS dataset and the simulation code on Github<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup>.