PINN: ripristina forward() originale, aumenta LBFGS_STEPS e corregge pesi loss

Il training collassava alla soluzione banale T=T_AMB perché W_BC=10 spingeva
Adam a soddisfare le Robin BC (trivialmente, con gradiente zero) sacrificando
la PDE. Fix: W_PDE=10, W_BC=1 così la PDE domina il gradiente fin dal primo
epoch. LBFGS_STEPS: 20→200 perché L-BFGS era l'unico ottimizzatore a fare
progressi reali. forward(): rimossa moltiplicazione t_norm che causava
vanishing gradient su dT/dx e d²T/dx².

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

model.py

@@ -14,14 +14,10 @@ class HeatPINN(nn.Module):
         self.net = nn.Sequential(*layers)
 
     def forward(self, xt):
-        # Normalize inputs to [0,1]x[0,1] so both dimensions share the same scale
-        x_norm = xt[:, 0:1] / config.L
-        t_norm = xt[:, 1:2] / config.T_END
-        xt_norm = torch.cat([x_norm, t_norm], dim=1)
-        # Multiply by t_norm to enforce IC as hard constraint: T(x,0) = T_AMB exactly.
-        # This eliminates the trivial solution (net=0) and forces non-trivial learning for t>0.
-        T_char = config.Q_VAL * config.L / config.K
-        return config.T_AMB + T_char * t_norm.squeeze(1) * self.net(xt_norm).squeeze(1)
+        # Output scaled to physical range: T_AMB + (Q*L/K) * net
+        # net learns dimensionless perturbation; inputs x∈[0,L], t∈[0,T_END] passed as-is
+        # so Tanh activations see t values up to T_END, providing implicit regularization
+        return config.T_AMB + (config.Q_VAL * config.L / config.K) * self.net(xt)
 
 
 # Precomputed loss scales (depend only on config constants)