TunerPINN, high Q makes problems

src/03_TunerPINN.py

@@ -8,6 +8,9 @@ import os
 
 from PINN import *
 from FEniCSx import solve_heatequation_dimensionless
+from ANN import load_data, prepare_data # for validation
+
+from sklearn.metrics import root_mean_squared_error
 
 import skopt
 from skopt import gp_minimize # Import the skopt library for hyperparameter optimization
@@ -27,7 +30,7 @@ beta = delta_T / T_ICE_lb  # β
 # Setup
 ####################################################################################################
 # HPO setting
-n_calls = 20
+n_calls = 15 #!20
 # Create solution folder for each iteration
 for i in range(n_calls):
     if not os.path.exists(f'models/PINN/TunerRuns/{i}'):
@@ -43,7 +46,7 @@ num_test = 2_000
 
 optimizer = 'adam'
 display_every = 1_000
-iterations = [3_000, 1_000]#![4_000, 4_000, 4_000]
+iterations = [3_000, 2_000]#![4_000, 4_000, 4_000]
 learning_rate = [1e-3, 1e-4]#![1e-3, 1e-4, 1e-5]
 
 # dim_num_dense_layers = Integer(low=1, high=9, name='num_dense_layers') 
@@ -67,6 +70,23 @@ default_parameters = [
 ]
 
 
+####################################################################################################
+# Validation data
+####################################################################################################
+data = load_data()
+data = prepare_data(data)
+time_vec = data['time_validate']
+Theta = data['Theta_ICE_validate']
+x_right = np.array([1.0])
+X, T_grid = np.meshgrid(x_right, time_vec)
+X_flat = X.flatten()
+T_flat = T_grid.flatten()
+samples = Theta.shape[0]
+X_input_PINN = []
+for i in range(samples):
+    Q_flat = np.full_like(X_flat, Theta[i], dtype=float)
+    X_input_PINN = np.vstack((X_flat, Q_flat, T_flat)).T
+X_input_PINN = np.array(X_input_PINN)
 
 ####################################################################################################
 # Hyperparameter Optimization
@@ -103,9 +123,9 @@ def fitness(num_dense_layers, num_dense_nodes, activation):
 
     # Plotting
     # Plot the Loss History
-    # dde.utils.plot_loss_history(losshistory)
-    # plt.savefig(f'models/PINN/TunerRuns/{ITERATION}//PINNTunerLossHistory.png')
-    # plt.savefig(f'models/PINN/TunerRuns/{ITERATION}//PINNTunerLossHistory.pdf')
+    dde.utils.plot_loss_history(losshistory)
+    plt.savefig(f'models/PINN/TunerRuns/{ITERATION}//PINNTunerLossHistory.png')
+    plt.savefig(f'models/PINN/TunerRuns/{ITERATION}//PINNTunerLossHistory.pdf')
 
     # Plot loss divided into components
     plot_loss_divided(losshistory, savefig=f'models/PINN/TunerRuns/{ITERATION}/PINNTunerLossHistory_splitted.pdf', show=False)
@@ -139,7 +159,11 @@ def fitness(num_dense_layers, num_dense_nodes, activation):
     plt.legend()
     plt.savefig(f'models/PINN/TunerRuns/{ITERATION}/HeatFluxRight.pdf')
 
-    error = np.array(losshistory.loss_train).sum(axis=1).ravel().min()
+    # Validation for heat flux
+    q_predict_PINN = np.array([heat_flux_right(model, Theta[i], time_vec) for i in range(samples)]).reshape(samples, -1)
+    rmse = root_mean_squared_error(data['q_right_validate'].T, q_predict_PINN.T, multioutput='raw_values')
+    error = np.median(rmse)
+    # error = np.array(losshistory.loss_train).sum(axis=1).ravel().min()
 
     print(f'Error: {error}')