PINNBench

PINNBench offers a platform to evaluate various architectures and methods for data-free PINN 1D designs, supporting an expanding set of PDE equations.

Installing

After cloning the project, follow these steps to install all dependencies:

Install Poetry by running:

pip install poetry

Build the project with:

poetry build

Running the code

To run the code, execute the main.py file, which contains all the controlling parameters.

To reproduce the results from the bachelor thesis, use the following configurations:

PINN = PINN_types.PINN_MLP.PINN(
    input_dim       = 2,                                                             
    output_dim      = 1,                                                              
    N_hidden        = 4,                                                               
    width_layer     = 200,                                                            
    domains         = [2*torch.pi,1],                                                          
    PDE_funct       = PINN_utils.Equations.PDE_equations.PDE_convection,                    
    Boundary_funct  = PINN_utils.Equations.Boundary_conditions.BC_equal,              
    Init_cond_funct = PINN_utils.Equations.Initial_conditions.IC_funct,              
    mapping         = None,                                                           
    act             = "wavelet"                                                          
)

As for configuring the main class,

main = Main(model = PINN.to(device),                  
                 epoch = 200000,                          
                 alpha_lr = 0.90,                     
                 lr_init = 1e-3,                      
                 min_lr  = 1e-9,                      
                 lr_freq = 5000,                     
                 weight_balance_freq = 2500,          
                 weight_balance     = "NTK",        
                 optimizer          = "Adam",         
                 training_scheme    = "classic",      
                 sampler_type       = "dynamic",       
                 sampler_domain     = [2*torch.pi,1], 
                 log  =True,                          
                 PDE_coefs = [90],                     
                 alpha = 0.9,                           
                 epsilon = 6e-6,                      
                 gradnorm_coefs=[0.5, 1000, 1e-4]     
            )