Authors:
- Nikolas Borrel-Jensen (nikolasborrel@gmail.com)
- Allan P. Engsig-Karup (apek@dtu.dk)
- Cheol-Ho Jeong (chje@dtu.dk)
Code for sound field predictions in domains with Neumann and impedance boundaries. Used for generating results from the paper "Physics-informed neural networks for 1D sound field predictions with parameterized sources and impedance boundaries" by N. Borrel-Jensen, A. P. Engsig-Karup, and C. Jeong.
This code is dependent on the SciANN library, which is not maintained anymore after version 0.7.0.1. SciANN is a wrapper around Tensorflow and Keras and the last test are based on python==3.9 and tensorflow==2.10. This means, that macs running on M2 silicon are not compatible with the code (requires tensorflow==2.13+).
You can install the required library versions used for reproducing the results from the paper with pip as (with --user argument if sudo access is not possible):
pip3 install --user tensorflow==2.5.1 sciann==0.6.4.7 python==3.8 matplotlib smt pydot graphviz
and/or using the ./requirements.txt file located in the root directory.
Run
python3 main_train.py --path_settings="path/to/script.json"Scripts for setting up models with Neumann, frequency-independent and dependent boundaries can be found in scripts/settings (see JSON settings).
Run
python3 main_evaluate.pyThe settings are
id_dir = <unique id>
settings_filename = 'settings.json'
base_dir = "path/to/base/dir"
do_plots_for_paper = <bool>
do_animations = <bool>
do_side_by_side_plot = <bool>The id_dir corresponds to the output directory generated after training, settings_filename is the name of the settings file used for training (located inside the id_dir directory), base_dir is the path to the base directory (see Input/output directory structure).
To evaluate the execution time of the surrogate model, run
python3 main_evaluate_timings.py --path_settings="path/to/script.json" --trained_model_tag="trained-model-dir"The trained_model_tag is the directory with the trained model weights trained using the scripts located at the path given in path_settings.
The library dependencies can be installed with pip and are
- tensorflow==2.5.1
- sciann==0.6.4.7
- matplotlib
- pandas
- smt
- pydot
- graphviz
See also scripts/package_install.sh.
The input data should be located inside a folder base_dir (can be given any name) with a specific relative directory structure as
base_dir/
reference_data/
freq_dep_1D_1000.00Hz_sigma0.2_c1_d0.02_srcs3.hdf5
...
freq_indep_1D_1000.00Hz_sigma0.2_c1_xi5.83_srcs3.hdf5
...
neumann_1D_1000.00Hz_sigma0.2_c1_srcs3.hdf5
...
...
The reference data from the paper can be downloaded here. The reference data is generated using an SEM solver for impedance boundaries, whereas the Python script main_generate_analytical_data.py was used for Neumann boundaries.
Note: the folder reference_data should be named exactly as stated.
Output result data will be located inside the results folder as
base_dir/
...
results/
id_folder/
figs/
models/
LossType.PINN/
checkpoint
cp.ckpt.data-00000-of-00001
cp.ckpt.index
settings.json
The settings.json file is identical to the settings file used for training indicated by the --path_settings argument. The directory LossType.PINN contains the trained model weights.
The results from the paper can be downloaded here here and should be placed inside the results directory (run python3 main_evaluate.py to evaluate).
For transfer learning, the model to continue the training on, should be located
base_path/
trained_models/
trained_model_tag/
checkpoint
cp.ckpt.data-00000-of-00001
cp.ckpt.index
...
where the directory trained_model_tag/ would correspond to the generated LossType.PINN/ model created after training (can be given any name/tag).
The script scripts/settings/neumann.json was used for training the Neumann model from the paper
{
"id": "neumann_srcs3_sine_3_256_7sources_loss02",
"base_path": "path/to/base_dir/",
"c": 1,
"c_phys": 343,
"___COMMENT_fmax___": "1000Hz*c/343 = 2.9155 for c=1",
"fmax": 2.9155,
"tmax": 4,
"Xmin": [-1],
"Xmax": [1],
"source_pos": [[-0.3],[-0.2],[-0.1],[0.0],[0.1],[0.2],[0.3]],
"sigma0": 0.2,
"rho": 1.2,
"ppw": 10,
"epochs": 25000,
"stop_loss_value": 0.0002,
"boundary_type": "NEUMANN",
"data_filename": "neumann_1D_1000.00Hz_sigma0.2_c1_srcs7.hdf5",
"batch_size": 512,
"learning_rate": 0.0001,
"optimizer": "adam",
"__comment0__": "NN setting for the PDE",
"activation": "sin",
"num_layers": 3,
"num_neurons": 256,
"ic_points_distr": 0.25,
"bc_points_distr": 0.45,
"loss_penalties": {
"pde":1,
"ic":20,
"bc":1
},
"verbose_out": false,
"show_plots": false
}The script scripts/settings/freq_indep.json was used for training the frequency-independent model from the paper
{
"id": "freq_indep_sine_3_256_7sources_loss02",
"base_path": "path/to/base_dir/",
"c": 1,
"c_phys": 343,
"___COMMENT_fmax___": "1000Hz*c/343 = 2.9155 for c=1",
"fmax": 2.9155,
"tmax": 4,
"Xmin": [-1],
"Xmax": [1],
"source_pos": [[-0.3],[-0.2],[-0.1],[0.0],[0.1],[0.2],[0.3]],
"sigma0": 0.2,
"rho": 1.2,
"ppw": 10,
"epochs": 25000,
"stop_loss_value": 0.0002,
"batch_size": 512,
"learning_rate": 0.0001,
"optimizer": "adam",
"boundary_type": "IMPEDANCE_FREQ_INDEP",
"data_filename": "freq_indep_1D_1000.00Hz_sigma0.2_c1_xi5.83_srcs7.hdf5",
"__comment0__": "NN setting for the PDE",
"activation": "sin",
"num_layers": 3,
"num_neurons": 256,
"impedance_data": {
"__comment1__": "xi is the acoustic impedance ONLY for freq. indep. boundaries",
"xi": 5.83
},
"ic_points_distr": 0.25,
"bc_points_distr": 0.45,
"loss_penalties": {
"pde":1,
"ic":20,
"bc":1
},
"verbose_out": false,
"show_plots": false
}The script scripts/settings/freq_dep.json was used for training the frequency-dependent model from the paper
{
"id": "freq_dep_sine_3_256_7sources_d01",
"base_path": "path/to/base_dir/",
"c": 1,
"c_phys": 343,
"___COMMENT_fmax___": "1000Hz*c/343 = 2.9155 for c=1",
"fmax": 2.9155,
"tmax": 4,
"Xmin": [-1],
"Xmax": [1],
"source_pos": [[-0.3],[-0.2],[-0.1],[0.0],[0.1],[0.2],[0.3]],
"sigma0": 0.2,
"rho": 1.2,
"ppw": 10,
"epochs": 50000,
"stop_loss_value": 0.0002,
"do_transfer_learning": false,
"boundary_type": "IMPEDANCE_FREQ_DEP",
"data_filename": "freq_dep_1D_1000.00Hz_sigma0.2_c1_d0.10_srcs7.hdf5",
"batch_size": 512,
"learning_rate": 0.0001,
"optimizer": "adam",
"__comment0__": "NN setting for the PDE",
"activation": "sin",
"num_layers": 3,
"num_neurons": 256,
"__comment1__": "NN setting for the auxillary differential ODE",
"activation_ade": "tanh",
"num_layers_ade": 3,
"num_neurons_ade": 20,
"impedance_data": {
"d": 0.1,
"type": "IMPEDANCE_FREQ_DEP",
"lambdas": [7.1109025021758407,205.64002739443146],
"alpha": [6.1969460587749818],
"beta": [-15.797795759219973],
"Yinf": 0.76935257750377573,
"A": [-7.7594660571346719,0.0096108036858666163],
"B": [-0.016951521199665469],
"C": [-2.4690553703530442]
},
"accumulator_factors": [10.26, 261.37, 45.88, 21.99],
"ic_points_distr": 0.25,
"bc_points_distr": 0.45,
"loss_penalties": {
"pde":1,
"ic":20,
"bc":1,
"ade":[10,10,10,10]
},
"verbose_out": false,
"show_plots": false
}The base_path should be adjusted to your local setup as described in section "Input/output directory structure".
The scripts for training the models on the GPULAB clusters at DTU are located at scripts/settings/run_*.sh.
Launch scripts for VS Code are located inside .vscode and running the settings script local_train.json in debug mode is done selecting the Python: TRAIN scheme (open pinn-acoustics.code-workspace to enable the workspace).
See LICENSE