diff --git a/pytest.ini b/pytest.ini
new file mode 100644
index 0000000..98f88c8
--- /dev/null
+++ b/pytest.ini
@@ -0,0 +1,5 @@
+[pytest]
+testpaths = tests
+addopts   = -v --tb=short
+markers =
+    slow: test lenti (training completo) — escludi con -m "not slow"
diff --git a/requirements.txt b/requirements.txt
index 687263c..e6f5b2f 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -1,4 +1,5 @@
 torch>=2.0.0
+pytest>=7.0.0
 pandas>=2.0.0
 numpy>=1.24.0
 scikit-learn>=1.3.0
diff --git a/tests/conftest.py b/tests/conftest.py
new file mode 100644
index 0000000..5593660
--- /dev/null
+++ b/tests/conftest.py
@@ -0,0 +1,24 @@
+import sys
+import os
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+import pytest
+import torch
+
+
+@pytest.fixture(scope="session")
+def device():
+    from engine import _get_device
+    return _get_device()
+
+
+@pytest.fixture
+def small_data():
+    from engine import prepare_data
+    return prepare_data(N_f=200, N_ic=50, N_bc=50)
+
+
+@pytest.fixture
+def pinn_model(device):
+    from model import HeatPINN
+    return HeatPINN().to(device)
diff --git a/tests/test_config.py b/tests/test_config.py
new file mode 100644
index 0000000..27d4634
--- /dev/null
+++ b/tests/test_config.py
@@ -0,0 +1,73 @@
+import sys
+import os
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+import math
+import config
+
+
+def test_x_src_within_domain():
+    assert 0.0 <= config.X_SRC <= config.L
+
+
+def test_t_step_before_t_end():
+    assert 0.0 < config.T_STEP < config.T_END
+
+
+def test_gauss_sigma_positive():
+    assert config.GAUSS_SIGMA > 0.0
+
+
+def test_physics_positive():
+    assert config.ALPHA > 0.0
+    assert config.K > 0.0
+    assert config.H_CONV > 0.0
+    assert config.L > 0.0
+    assert config.T_END > 0.0
+
+
+def test_training_hyperparameters_positive():
+    assert config.PATIENCE > 0
+    assert config.EPOCHS > 0
+    assert config.LR_ADAM > 0.0
+    assert config.SCHED_MIN_LR > 0.0
+    assert config.SCHED_FACTOR > 0.0
+    assert config.SCHED_PATIENCE > 0
+
+
+def test_lr_ordering():
+    """Min LR deve essere inferiore all'LR iniziale."""
+    assert config.SCHED_MIN_LR < config.LR_ADAM
+
+
+def test_sched_patience_lt_patience():
+    """Lo scheduler deve poter agire prima che scatti l'early stopping."""
+    assert config.SCHED_PATIENCE < config.PATIENCE
+
+
+def test_cfl_stability():
+    """La griglia FDM deve soddisfare la condizione CFL (r ≤ 0.5)."""
+    dx = config.L / (config.NX - 1)
+    dt = config.T_END / (config.NT - 1)
+    r = config.ALPHA * dt / dx ** 2
+    assert r <= 0.5, f"CFL violata: r={r:.4f} > 0.5"
+
+
+def test_grid_dimensions():
+    assert config.NX >= 2
+    assert config.NT >= 2
+    assert config.N_F >= 1
+    assert config.N_IC >= 1
+    assert config.N_BC >= 1
+
+
+def test_pde_scale_covers_source_peak():
+    """_pde_scale in model.py deve coprire il picco gaussiano della sorgente."""
+    from model import _pde_scale
+    src_peak = config.ALPHA * config.Q_VAL / (
+        config.K * config.GAUSS_SIGMA * math.sqrt(2 * math.pi)
+    )
+    assert _pde_scale >= src_peak ** 2 - 1e-6, (
+        f"_pde_scale={_pde_scale:.1f} < src_peak²={src_peak**2:.1f}: "
+        "la loss PDE non è normalizzata correttamente"
+    )
diff --git a/tests/test_e2e.py b/tests/test_e2e.py
new file mode 100644
index 0000000..e086073
--- /dev/null
+++ b/tests/test_e2e.py
@@ -0,0 +1,87 @@
+import sys
+import os
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+import pytest
+import numpy as np
+import torch
+import config
+
+
+# ── FDM ───────────────────────────────────────────────────────────────────────
+
+def test_fdm_full_run():
+    """Il solver FDM produce un campo di temperatura fisicamente corretto."""
+    from fdm.solver import solve
+    T, x, t = solve()
+
+    assert T.shape == (config.NX, config.NT)
+    assert np.isfinite(T).all()
+    assert T[:, -1].mean() > config.T0          # la sorgente ha scaldato il dominio
+    assert T.max() > config.T0                   # picco sopra la IC
+    assert T.min() >= config.T0 - 1e-6          # nessun raffreddamento sotto T0 (no sorgenti fredde)
+
+
+def test_fdm_visualizer_creates_html(tmp_path, monkeypatch):
+    """Il visualizer FDM scrive almeno un file HTML senza errori."""
+    import fdm.visualizer as fdm_vis
+    monkeypatch.setattr(fdm_vis, 'BASE_DIR', str(tmp_path))
+
+    from fdm.solver import solve
+    T, x, t = solve()
+    fdm_vis.visualize_fdm(T, x, t)
+
+    html_files = list(tmp_path.rglob('*.html'))
+    assert len(html_files) >= 1, "Nessun file HTML generato dal visualizer FDM"
+
+
+def test_pinn_visualizer_creates_html(tmp_path, monkeypatch):
+    """Il visualizer PINN scrive i tre file HTML senza errori."""
+    import visualizer as pinn_vis
+    monkeypatch.setattr(pinn_vis, 'BASE_DIR', str(tmp_path))
+
+    from fdm.solver import solve as fdm_solve
+    T_fdm, _, _ = fdm_solve()
+
+    nx, nt = 20, 20
+    x_vals = np.linspace(0, config.L, nx)
+    t_vals = np.linspace(0, config.T_END, nt)
+    T_pred = np.full((nx, nt), config.T_AMB)    # predizione costante (dummy)
+
+    pinn_vis.visualize_heat_field(T_pred, x_vals, t_vals, T_fdm)
+
+    html_files = list(tmp_path.rglob('*.html'))
+    assert len(html_files) == 3, f"Attesi 3 HTML, trovati {len(html_files)}"
+
+
+# ── PINN training (lento) ─────────────────────────────────────────────────────
+
+@pytest.mark.slow
+def test_pinn_training_saves_checkpoint(tmp_path, monkeypatch):
+    """Training per 30 epoche: il checkpoint viene salvato."""
+    import engine
+    save_path = str(tmp_path / 'model.pth')
+    monkeypatch.setattr(engine, 'MODEL_SAVE_PATH', save_path)
+    monkeypatch.setattr(engine, 'MODELS_DIR', str(tmp_path))
+
+    from engine import prepare_data, train_model
+    data = prepare_data(N_f=300, N_ic=100, N_bc=100)
+    train_model(data, epochs=30, patience=30)
+
+    assert os.path.exists(save_path)
+    ckpt = torch.load(save_path, map_location='cpu', weights_only=True)
+    assert 'state_dict' in ckpt
+
+
+@pytest.mark.slow
+def test_pinn_evaluate_after_training(tmp_path, monkeypatch):
+    """evaluate_model gira senza errori dopo un training minimo."""
+    import engine
+    save_path = str(tmp_path / 'model.pth')
+    monkeypatch.setattr(engine, 'MODEL_SAVE_PATH', save_path)
+    monkeypatch.setattr(engine, 'MODELS_DIR', str(tmp_path))
+
+    from engine import prepare_data, train_model, evaluate_model
+    data = prepare_data(N_f=300, N_ic=100, N_bc=100)
+    train_model(data, epochs=30, patience=30)
+    evaluate_model(data)
diff --git a/tests/test_engine_data.py b/tests/test_engine_data.py
new file mode 100644
index 0000000..5737297
--- /dev/null
+++ b/tests/test_engine_data.py
@@ -0,0 +1,67 @@
+import sys
+import os
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+import torch
+import config
+from engine import set_seed, _get_device, prepare_data
+
+
+def test_set_seed_reproducibility():
+    set_seed(42)
+    r1 = torch.rand(10)
+    set_seed(42)
+    r2 = torch.rand(10)
+    torch.testing.assert_close(r1, r2)
+
+
+def test_get_device_valid():
+    device = _get_device()
+    assert isinstance(device, torch.device)
+    assert device.type in ('cpu', 'cuda', 'mps')
+
+
+def test_prepare_data_keys():
+    data = prepare_data(N_f=100, N_ic=50, N_bc=50)
+    assert set(data.keys()) == {'device', 'x_f', 't_f', 'x_ic', 't_bc'}
+
+
+def test_prepare_data_shapes():
+    N_f, N_ic, N_bc = 100, 50, 50
+    data = prepare_data(N_f=N_f, N_ic=N_ic, N_bc=N_bc)
+    # engine.py aggiunge 2 * (N_f // 4) punti di clustering
+    expected_f = N_f + 2 * (N_f // 4)
+    assert data['x_f'].shape == (expected_f,)
+    assert data['t_f'].shape == (expected_f,)
+    assert data['x_ic'].shape == (N_ic,)
+    assert data['t_bc'].shape == (N_bc,)
+
+
+def test_prepare_data_x_bounds():
+    data = prepare_data(N_f=500, N_ic=100, N_bc=100)
+    assert data['x_f'].min().item() >= 0.0
+    assert data['x_f'].max().item() <= config.L
+    assert data['x_ic'].min().item() >= 0.0
+    assert data['x_ic'].max().item() <= config.L
+
+
+def test_prepare_data_t_bounds():
+    data = prepare_data(N_f=500, N_ic=100, N_bc=100)
+    assert data['t_f'].min().item() >= 0.0
+    assert data['t_f'].max().item() <= config.T_END
+
+
+def test_prepare_data_device_consistency():
+    data = prepare_data(N_f=100, N_ic=50, N_bc=50)
+    expected = data['device'].type
+    for key in ('x_f', 't_f', 'x_ic', 't_bc'):
+        assert data[key].device.type == expected, f"{key} sul device sbagliato"
+
+
+def test_prepare_data_deterministic():
+    """Due chiamate con lo stesso seed (fissato in prepare_data) producono dati identici."""
+    d1 = prepare_data(N_f=100, N_ic=50, N_bc=50)
+    d2 = prepare_data(N_f=100, N_ic=50, N_bc=50)
+    torch.testing.assert_close(d1['x_f'], d2['x_f'])
+    torch.testing.assert_close(d1['t_f'], d2['t_f'])
+    torch.testing.assert_close(d1['x_ic'], d2['x_ic'])
diff --git a/tests/test_integration_pinn.py b/tests/test_integration_pinn.py
new file mode 100644
index 0000000..ee9267e
--- /dev/null
+++ b/tests/test_integration_pinn.py
@@ -0,0 +1,65 @@
+import sys
+import os
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+import torch
+from engine import prepare_data
+from model import HeatPINN, heat_pinn_loss
+
+
+def test_data_to_model_forward():
+    """prepare_data → forward: shape e device coerenti, nessun NaN."""
+    data = prepare_data(N_f=100, N_ic=50, N_bc=50)
+    model = HeatPINN().to(data['device'])
+    xt = torch.stack([data['x_f'], data['t_f']], dim=1)
+    out = model(xt)
+    assert out.shape == (data['x_f'].shape[0], 1)
+    assert out.device.type == data['device'].type
+    assert torch.isfinite(out).all()
+
+
+def test_full_loss_pipeline():
+    """prepare_data → heat_pinn_loss: tutti i componenti finiti e non-negativi."""
+    data = prepare_data(N_f=100, N_ic=50, N_bc=50)
+    model = HeatPINN().to(data['device'])
+    total, L_pde, L_ic, L_bc = heat_pinn_loss(
+        model, data['x_f'], data['t_f'], data['x_ic'], data['t_bc']
+    )
+    for name, v in [('total', total), ('L_pde', L_pde), ('L_ic', L_ic), ('L_bc', L_bc)]:
+        assert torch.isfinite(v), f"{name} non è finita"
+        assert v.item() >= 0.0, f"{name} è negativa"
+
+
+def test_backward_gradients_finite():
+    """Il backward della loss non produce NaN/Inf nei parametri."""
+    data = prepare_data(N_f=100, N_ic=50, N_bc=50)
+    model = HeatPINN().to(data['device'])
+    optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)
+    optimizer.zero_grad()
+    total, _, _, _ = heat_pinn_loss(
+        model, data['x_f'], data['t_f'], data['x_ic'], data['t_bc']
+    )
+    total.backward()
+    for p in model.parameters():
+        assert p.grad is not None
+        assert torch.isfinite(p.grad).all(), "gradiente NaN/Inf"
+
+
+def test_training_loop_stable():
+    """20 step di Adam non producono NaN/Inf nei parametri né nella loss."""
+    data = prepare_data(N_f=200, N_ic=100, N_bc=100)
+    model = HeatPINN().to(data['device'])
+    optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)
+    args = (model, data['x_f'], data['t_f'], data['x_ic'], data['t_bc'])
+
+    for _ in range(20):
+        optimizer.zero_grad()
+        loss, _, _, _ = heat_pinn_loss(*args)
+        loss.backward()
+        optimizer.step()
+
+    for p in model.parameters():
+        assert torch.isfinite(p).all(), "parametro NaN/Inf dopo training"
+
+    total, _, _, _ = heat_pinn_loss(*args)
+    assert torch.isfinite(total)
diff --git a/tests/test_model.py b/tests/test_model.py
new file mode 100644
index 0000000..0ca83e3
--- /dev/null
+++ b/tests/test_model.py
@@ -0,0 +1,102 @@
+import sys
+import os
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+import torch
+import config
+from model import HeatPINN, heat_pinn_loss
+
+
+# ── HeatPINN.forward ─────────────────────────────────────────────────────────
+
+def test_forward_output_shape(pinn_model, device):
+    xt = torch.zeros(64, 2, device=device)
+    xt[:, 0] = torch.rand(64) * config.L
+    xt[:, 1] = torch.rand(64) * config.T_END
+    assert pinn_model(xt).shape == (64, 1)
+
+
+def test_forward_finite(pinn_model, device):
+    xt = torch.zeros(100, 2, device=device)
+    xt[:, 0] = torch.rand(100) * config.L
+    xt[:, 1] = torch.rand(100) * config.T_END
+    assert torch.isfinite(pinn_model(xt)).all()
+
+
+def test_forward_zero_weights_returns_t_amb(device):
+    """Con pesi nulli net(x,t)=0 ⇒ forward restituisce T_AMB per ogni input."""
+    model = HeatPINN().to(device)
+    for p in model.parameters():
+        p.data.zero_()
+    xt = torch.zeros(20, 2, device=device)
+    xt[:, 0] = torch.rand(20) * config.L
+    xt[:, 1] = torch.rand(20) * config.T_END
+    out = model(xt)
+    torch.testing.assert_close(out, torch.full_like(out, config.T_AMB), atol=1e-5, rtol=0.0)
+
+
+def test_forward_t_normalization(device):
+    """t viene normalizzato a [0,1]: il modello deve restituire output finiti
+    anche a t=T_END senza saturazione di Tanh."""
+    model = HeatPINN().to(device)
+    torch.nn.init.xavier_uniform_(model.net[0].weight)
+    xt = torch.tensor([[0.5, 0.0], [0.5, config.T_END]], device=device)
+    out = model(xt)
+    assert out.shape == (2, 1)
+    assert torch.isfinite(out).all()
+
+
+# ── heat_pinn_loss ────────────────────────────────────────────────────────────
+
+def _dummy_inputs(device, n_f=100, n_ic=50, n_bc=50):
+    x_f  = torch.rand(n_f,  device=device) * config.L
+    t_f  = torch.rand(n_f,  device=device) * config.T_END
+    x_ic = torch.rand(n_ic, device=device) * config.L
+    t_bc = torch.rand(n_bc, device=device) * config.T_END
+    return x_f, t_f, x_ic, t_bc
+
+
+def test_loss_returns_four_values(pinn_model, device):
+    result = heat_pinn_loss(pinn_model, *_dummy_inputs(device))
+    assert len(result) == 4
+
+
+def test_loss_components_non_negative(pinn_model, device):
+    total, L_pde, L_ic, L_bc = heat_pinn_loss(pinn_model, *_dummy_inputs(device))
+    assert total.item() >= 0.0
+    assert L_pde.item() >= 0.0
+    assert L_ic.item() >= 0.0
+    assert L_bc.item() >= 0.0
+
+
+def test_loss_finite(pinn_model, device):
+    for v in heat_pinn_loss(pinn_model, *_dummy_inputs(device)):
+        assert torch.isfinite(v), f"loss non finita: {v}"
+
+
+def test_loss_weight_doubles_pde_contribution(pinn_model, device):
+    """Raddoppiare w_pde con w_ic=w_bc=0 deve raddoppiare il totale."""
+    inputs = _dummy_inputs(device)
+    total1, L_pde1, _, _ = heat_pinn_loss(pinn_model, *inputs, w_pde=1.0, w_ic=0.0, w_bc=0.0)
+    total2, L_pde2, _, _ = heat_pinn_loss(pinn_model, *inputs, w_pde=2.0, w_ic=0.0, w_bc=0.0)
+    # L_pde deve essere identico tra le due chiamate (stesso modello, stessi dati)
+    torch.testing.assert_close(L_pde1, L_pde2, atol=1e-5, rtol=1e-4)
+    torch.testing.assert_close(total2, 2.0 * total1, atol=1e-5, rtol=1e-4)
+
+
+def test_ic_loss_zero_when_net_is_zero(device):
+    """Con net=0 ⇒ T = T_AMB = T0 ⇒ L_ic = 0."""
+    model = HeatPINN().to(device)
+    for p in model.parameters():
+        p.data.zero_()
+    _, _, L_ic, _ = heat_pinn_loss(model, *_dummy_inputs(device))
+    assert L_ic.item() < 1e-8
+
+
+def test_bc_loss_zero_when_net_is_zero(device):
+    """Con net=0 ⇒ T = T_AMB e dT/dx = 0 ⇒ Robin BC soddisfatta ⇒ L_bc = 0."""
+    model = HeatPINN().to(device)
+    for p in model.parameters():
+        p.data.zero_()
+    _, _, _, L_bc = heat_pinn_loss(model, *_dummy_inputs(device))
+    assert L_bc.item() < 1e-8