diff --git a/surrogate_handler.py b/surrogate_handler.py
new file mode 100644
index 0000000..20d3bb5
--- /dev/null
+++ b/surrogate_handler.py
@@ -0,0 +1,40 @@
+import numpy as np
+from sklearn.neural_network import MLPRegressor
+from sklearn.ensemble import RandomForestRegressor
+
+class SurrogateHandler:
+    def __init__(self, model_type='mlp'):
+        self.model_type = model_type
+        self.is_trained = False
+        self.data_X = []
+        self.data_Y = []
+        
+        # Model choice
+        if model_type == 'mlp':
+            self.model = MLPRegressor(hidden_layer_sizes=(100, 50), max_iter=500, random_state=42)
+        elif model_type == 'rf':
+            # RandomForest is generaly more robust "out of the box"
+            self.model = RandomForestRegressor(n_estimators=100, random_state=42)
+        else:
+            raise ValueError("Model type must be 'mlp' or 'rf'")
+
+    def add_data(self, x_matrix, f2_value):
+        # Flattening the position matrix to a 1 dimension vector
+        flat_x = np.array(x_matrix).flatten()
+        self.data_X.append(flat_x)
+        self.data_Y.append(f2_value)
+
+    def train(self):
+        if len(self.data_X) < 20: # No training if their is too few data
+            return
+        
+        X = np.array(self.data_X)
+        y = np.array(self.data_Y)
+        self.model.fit(X, y)
+        self.is_trained = True
+
+    def predict(self, x_matrix):
+        if not self.is_trained:
+            return None
+        flat_x = np.array(x_matrix).flatten().reshape(1, -1)
+        return self.model.predict(flat_x)[0]