Physical Modeling using Recurrent Neural Networks with Fast Convolutional Layers

This is the accompanying repository to the paper submission for DAFx20in22, Vienna, Austria by Julian D. Parker, Sebastian J. Schlecht, Rudolf Rabenstein, Maximilian Schäfer. The paper is currently in review.

View the Project on GitHub julian-parker/DAFX22_FNO

Supplemental Material for Physical Modeling using Recurrent Neural Networks with Fast Convolutional Layers

Example 1 - Linear String

Linear string GRU

Linear string Real

Linear string Reference

Linear string RNN

Example 2 - Nonlinear String

Nonlinear string GRU

Nonlinear string Real

Nonlinear string Reference

Nonlinear string RNN

Example 3 - 2D Wave Equation

Physical parameters used for dataset generation

Linear String

Quantity Value
\(E\) 5.4 e9
$\rho_s$ 1140
$l$ 0.65
$A$ 0.5188 e-6
$I$ 0.171 e-12
$d_1$ 8 e-1
$d_3$ 1.4 e-5
$T_s$ 60.97
$\nu$ 50

2d wave equation

Quantity Value
$l_x$ 1
$l_y$ 0.95
$c_0$ 340
$\rho_o$ 1.2041

Nonlinear Tension Modulated String

Quantity Value
$l$ 0.65
$A$ 0.5188e-6
$I$ 0.171e-12
$\rho$ 1140
$E$ 5.4e9
$d_1$ 1e-2
$d_3$ 6e-5
$T_{s0}$ 60.97

Network/Training Hyperparameters

Visible in: train_1d_string.py train_2d_wave.py
train_1d_nonlinear_string.py