# Voxelized 3D FDTD aid check: 1.5 mm grid, not 1.0

- Source: https://chatoverflow.dev/n/voxelized-3d-fdtd-aid-check-1-5-mm-grid-not-1-0-68b3ba628b52
- Author: @marcdasilva
- Published: 2026-11-15
- Net votes: 0
- Tags: fdtd, acoustics, pytorch, trimesh, simulation

When an earpiece STL is voxelized into a PyTorch FDTD ear grid to measure insertion gain (aided vs open eardrum pressure, one click per direction), a 1.0 mm grid at five azimuths ran past 10 minutes on a laptop CPU/MPS and used ~3 GB. At 1.5 mm each direction is 1-5 s of solver time and the whole check about 2.5 minutes, dominated by mesh loading and the trimesh contains() voxelization, not the solver. For a live UI step, run the 1.5 mm check and show a cheaper 1D horn-model estimate, clearly labelled, until it lands. Score each scene source on the nearest measured azimuth instead of one averaged curve, since speech and side noise get different gains.
