[Iter 328] Code modification in PDE_D_DeadzoneCTC.py

[Automated commit by Claude]

Author: allierc

src/ParticleGraph/generators/PDE_D_DeadzoneCTC.py

@@ -0,0 +1,362 @@
+import torch
+import torch_geometric as pyg
+import torch_geometric.utils as pyg_utils
+from ParticleGraph.utils import to_numpy
+
+class PDE_D_DeadzoneCTC(pyg.nn.MessagePassing):
+    """
+    CTC with receptor-adaptation deadzone + pp damping + fp drag.
+
+    Modifies the standard CTC tanh transfer function by introducing a flat
+    "deadzone" of width epsilon around the threshold concentration T. When
+    |C1 - T| < epsilon*A, the CTC sign factor is exactly zero, meaning
+    particles near the threshold experience NO chemotactic driving force.
+    They settle purely through pp damping. Outside the deadzone, the standard
+    tanh response applies with shifted argument.
+
+    This addresses the 2-type convergence bottleneck: particles near T
+    oscillate because the tanh function transitions smoothly through zero,
+    producing small but sign-alternating forces as C1 fluctuates around T.
+    The deadzone eliminates these oscillations entirely within the
+    adaptation zone.
+
+    Physical motivation: receptor adaptation / perfect adaptation in
+    chemotaxis. Cells near their target concentration adapt their receptor
+    sensitivity, becoming insensitive to small fluctuations. Only large
+    deviations from the adapted state trigger a migration response.
+
+    Transfer function:
+        sign_factor = 0                                        if |C1-T| < eps*A
+        sign_factor = -tanh(steep * (C1-T-eps*A) / A)         if C1 > T + eps*A
+        sign_factor = -tanh(steep * (C1-T+eps*A) / A)         if C1 < T - eps*A
+
+    Also includes all DragDampedCTC features: durotaxis, pp damping, fp drag.
+
+    Literature:
+    - Barkai, N. & Leibler, S. (1997) Nature 387:913-917
+      "Robustness in simple biochemical networks"
+    - Wolpert, L. (1969) J Theor Biol 25:1-47
+      "Positional information and the spatial pattern of cellular differentiation"
+    - Lo, C. M. et al. (2000) Biophysical Journal 79:144-152
+    - Painter, K. J. & Hillen, T. (2002) Can Appl Math Q 10(4):501-543
+    - Tranquillo, R. T. & Lauffenburger, D. A. (1987) J Math Biol 25:229-262
+
+    Per-type params layout: [M1, M2, consumption, production, ar_p1, ar_p2, ar_p3, ar_p4]
+    """
+
+    PARAMS_DOC = {
+        "model_name": "DeadzoneCTC",
+        "literature": "Barkai & Leibler (1997) Nature 387:913; Wolpert (1969); Lo (2000); Painter & Hillen (2002); Tranquillo (1987)",
+        "description": "CTC with receptor-adaptation deadzone near threshold + pp damping + fp drag",
+        "equations": {
+            "field_to_particle": "v = M*(1+alpha*|gradC1|)*sign_factor*grad*dir / (1+fp_drag*|vel|/v_ref)",
+            "deadzone_ctc": "sign_factor = 0 if |C1-T|<eps*A, else -tanh(3*(C1-T∓eps*A)/A)",
+            "particle_to_field": "dC1 = -consumption * w(r), dC2 = production * w(r)",
+            "particle_to_particle": "f = f_AR * (1 - damping * exp(-(C1_i - T)^2 / (2*width^2)))"
+        },
+        "params_mesh": [
+            {
+                "row": 0, "description": "C1 field parameters + CTC threshold",
+                "slots": [
+                    {"index": 0, "name": "D1", "description": "Diffusion coeff for C1"},
+                    {"index": 1, "name": "Da_c", "description": "Damkohler number"},
+                    {"index": 2, "name": "A", "description": "Brusselator A"},
+                    {"index": 3, "name": "B", "description": "Brusselator B"},
+                    {"index": 4, "name": "mu", "description": "Morphological param"},
+                    {"index": 5, "name": "M1", "description": "Mobility for C1 gradients"},
+                    {"index": 6, "name": "grad_amp_alpha", "description": "Durotaxis amplification"},
+                    {"index": 7, "name": "ctc_threshold", "description": "CTC threshold (T=ctc*A)"}
+                ]
+            },
+            {
+                "row": 1, "description": "C2 field + pp damping + deadzone params",
+                "slots": [
+                    {"index": 0, "name": "D2", "description": "Diffusion coeff for C2"},
+                    {"index": 1, "name": "M2", "description": "Mobility for C2 gradients"},
+                    {"index": 2, "name": "pp_damping", "description": "pp damping strength near T"},
+                    {"index": 3, "name": "pp_damping_width", "description": "Width of pp damping zone"},
+                    {"index": 4, "name": "deadzone_eps", "description": "Deadzone half-width (units of A). 0=standard CTC."},
+                    {"index": 5, "name": "unused1", "description": "Unused (pad)"}
+                ]
+            },
+            {
+                "row": 2, "description": "Particle-field coupling + fp drag",
+                "slots": [
+                    {"index": 0, "name": "Pe", "description": "Peclet number"},
+                    {"index": 1, "name": "consumption", "description": "Consumption rate of C1"},
+                    {"index": 2, "name": "production", "description": "Production rate of C2"},
+                    {"index": 3, "name": "influence_radius", "description": "Gaussian pf influence radius"},
+                    {"index": 4, "name": "fp_drag", "description": "Velocity-dependent fp drag"},
+                    {"index": 5, "name": "cross_type_factor", "description": "Per-type CTC threshold spread"}
+                ]
+            }
+        ],
+        "width_constraint": "ALL rows of params_mesh MUST have same number of columns (8). Pad shorter rows with 0.0."
+    }
+
+    def __init__(self, aggr_type='mean', p=None, particle_params=None, bc_dpos=None, dimension=2, sigma=0.005):
+        super(PDE_D_DeadzoneCTC, self).__init__(aggr=aggr_type)
+
+        self.p = p
+        self.particle_params = particle_params
+        self.bc_dpos = bc_dpos
+        self.dimension = dimension
+        self.sigma = sigma
+
+        self.M1 = p[0, 5]
+        self.M2 = p[1, 1]
+        self.consumption_rate = p[2, 1]
+        self.production_rate = p[2, 2]
+        self.influence_radius = p[2, 3]
+        self.Pe = p[2, 0]
+        self.repulsion_strength = 50
+        self.repulsion_range = 0.04
+
+        # Durotaxis gradient amplification
+        self.grad_amp_alpha = p[0, 6] if p.shape[1] > 6 else 0.0
+
+        # CTC threshold
+        self.ctc_threshold = p[0, 7] if p.shape[1] > 7 else 0.0
+        self.A_ref = p[0, 2]
+
+        # Per-type threshold spread
+        self.cross_type_factor = p[2, 5] if p.shape[1] > 5 else 0.0
+
+        # pp damping parameters (Painter & Hillen 2002)
+        self.pp_damping = p[1, 2] if p.shape[1] > 2 else 0.0
+        self.pp_damping_width = p[1, 3] if p.shape[1] > 3 else 0.5
+
+        # Deadzone half-width (Barkai & Leibler 1997)
+        self.deadzone_eps = p[1, 4] if p.shape[1] > 4 else 0.0
+
+        # Velocity-dependent fp drag (Tranquillo & Lauffenburger 1987)
+        self.fp_drag = p[2, 4] if p.shape[1] > 4 else 0.0
+        self.v_ref = 0.01
+
+        print(f"initialized PDE_D_DeadzoneCTC with parameters:")
+        print(f"  mobility: M1={self.M1.item()}, M2={self.M2.item()}")
+        ga_val = self.grad_amp_alpha.item() if hasattr(self.grad_amp_alpha, 'item') else self.grad_amp_alpha
+        print(f"  grad_amp_alpha={ga_val:.3f} (durotaxis, Lo 2000)")
+        ctc_val = self.ctc_threshold.item() if hasattr(self.ctc_threshold, 'item') else self.ctc_threshold
+        T_val = ctc_val * self.A_ref.item()
+        print(f"  ctc_threshold={ctc_val:.3f} (T={T_val:.2f}, Wolpert 1969)")
+        eps_val = self.deadzone_eps.item() if hasattr(self.deadzone_eps, 'item') else self.deadzone_eps
+        print(f"  DEADZONE: eps={eps_val:.3f} (half-width={eps_val*self.A_ref.item():.3f}, Barkai & Leibler 1997)")
+        print(f"    Particles within |C1-T| < {eps_val*self.A_ref.item():.3f} experience ZERO fp force")
+        damp_val = self.pp_damping.item() if hasattr(self.pp_damping, 'item') else self.pp_damping
+        damp_w = self.pp_damping_width.item() if hasattr(self.pp_damping_width, 'item') else self.pp_damping_width
+        print(f"  pp_damping={damp_val:.3f}, pp_damping_width={damp_w:.3f} (Painter & Hillen 2002)")
+        fp_drag_val = self.fp_drag.item() if hasattr(self.fp_drag, 'item') else self.fp_drag
+        print(f"  fp_drag={fp_drag_val:.3f}, v_ref={self.v_ref:.4f} (Tranquillo 1987)")
+        ctf_val = self.cross_type_factor.item() if hasattr(self.cross_type_factor, 'item') else self.cross_type_factor
+        if ctf_val > 0 and particle_params is not None:
+            n_types = particle_params.shape[0]
+            mean_idx = (n_types - 1) / 2.0
+            for t in range(n_types):
+                t_offset = ctf_val * (t - mean_idx)
+                t_val_t = T_val * (1.0 + t_offset)
+                print(f"    Type {t}: CTC threshold = {t_val_t:.2f} (offset={t_offset:+.2f})")
+        print(f"  Pe={self.Pe.item():.3f}, sigma={self.sigma}")
+        print(f"  particle->field: consumption={self.consumption_rate.item()}, production={self.production_rate.item()}, influence_radius={self.influence_radius.item():.3f}")
+        if particle_params is not None:
+            print(f"  multi-type support: {particle_params.shape[0]} particle types")
+
+    def forward(self, data, direction='fp'):
+        x, edge_index = data.x, data.edge_index
+        edge_index, _ = pyg_utils.remove_self_loops(edge_index)
+
+        if self.particle_params is not None:
+            particle_type = x[:, 1 + 2*self.dimension].long()
+            max_type = particle_type.max().item()
+            n_param_rows = self.particle_params.shape[0]
+            if max_type >= n_param_rows:
+                raise ValueError(
+                    f"PDE_D_DeadzoneCTC: particle_params has {n_param_rows} rows but found "
+                    f"particle type {max_type}. Need {max_type + 1} rows in simulation.params."
+                )
+            parameters = self.particle_params[to_numpy(particle_type), :]
+        else:
+            parameters = None
+
+        if direction == 'interpolate':
+            result = self.propagate(edge_index, x=x, mode='interpolate', parameters=parameters)
+            pos = x[:, 1:self.dimension+1]
+            in_box = ((pos >= 0) & (pos <= 1)).all(dim=1, keepdim=True)
+            result = result * in_box.float()
+            return result
+        elif direction == 'fp':
+            result = self.propagate(edge_index, x=x, mode='fp', parameters=parameters)
+            pos = x[:, 1:self.dimension+1]
+            in_box = ((pos >= 0) & (pos <= 1)).all(dim=1, keepdim=True)
+            result = result * in_box.float()
+            return result
+        elif direction == 'pf':
+            result = self.propagate(edge_index, x=x, mode='pf', parameters=parameters)
+            return result
+        else:
+            result = self.propagate(edge_index, x=x, mode='pp', parameters=parameters)
+            return result
+
+    def message(self, edge_index_i, edge_index_j, x_i, x_j, mode=None, parameters_i=None):
+        pos_i = x_i[:, 1:self.dimension+1]
+        pos_j = x_j[:, 1:self.dimension+1]
+
+        d_pos = self.bc_dpos(pos_j - pos_i)
+        dist = torch.sqrt(torch.sum(d_pos**2, dim=1))
+        dist_safe = torch.clamp(dist, min=1e-6)
+
+        if mode == 'interpolate':
+            C1_mesh = x_j[:, 6:7]
+            C2_mesh = x_j[:, 7:8]
+            weight = torch.exp(-dist / 0.01).unsqueeze(1)
+            return torch.cat([C1_mesh * weight, C2_mesh * weight, weight], dim=1)
+
+        elif mode == 'fp':
+            fields_i = x_i[:, 6:8]
+            fields_j = x_j[:, 6:8]
+
+            dC1 = fields_j[:, 0:1] - fields_i[:, 0:1]
+            dC2 = fields_j[:, 1:2] - fields_i[:, 1:2]
+
+            kernel = torch.exp(-dist / 0.05)
+            dir_norm = d_pos / dist_safe.unsqueeze(1)
+            domain_scale = 32.0
+            grad_C1 = (dC1 * kernel.unsqueeze(1)) / (dist_safe.unsqueeze(1) * domain_scale)
+            grad_C2 = (dC2 * kernel.unsqueeze(1)) / (dist_safe.unsqueeze(1) * domain_scale)
+
+            if parameters_i is not None:
+                M1 = parameters_i[:, 0:1]
+                M2 = parameters_i[:, 1:2]
+            else:
+                M1 = self.M1
+                M2 = self.M2
+
+            velocity_raw = (M1 * grad_C1 + M2 * grad_C2) * dir_norm
+
+            # 1. Durotaxis: amplify velocity at steep gradients (Lo et al. 2000)
+            if self.grad_amp_alpha > 0:
+                grad_mag = torch.abs(grad_C1)
+                grad_mag_clamped = torch.clamp(grad_mag, max=1.0)
+                amp_factor = 1.0 + self.grad_amp_alpha * grad_mag_clamped
+                velocity_raw = velocity_raw * amp_factor
+
+            # 2. CTC with deadzone (Wolpert 1969 + Barkai & Leibler 1997)
+            if self.ctc_threshold > 0:
+                C1_local = fields_i[:, 0:1]
+                A_ref = self.A_ref
+                base_T = self.ctc_threshold * A_ref
+                steepness = 3.0
+
+                # Per-type thresholds
+                if (parameters_i is not None and self.cross_type_factor > 0
+                        and x_i.numel() > 0):
+                    type_i = x_i[:, 1 + 2*self.dimension].long()
+                    n_types = type_i.max().item() + 1 if type_i.numel() > 0 else 1
+                    mean_idx = (n_types - 1) / 2.0
+                    type_offset = self.cross_type_factor * (type_i.float() - mean_idx)
+                    T = base_T * (1.0 + type_offset.unsqueeze(1))
+                else:
+                    T = base_T
+
+                deviation = C1_local - T
+                eps_width = self.deadzone_eps * A_ref  # deadzone half-width in C1 units
+
+                if eps_width > 0:
+                    # Deadzone CTC: zero force within |deviation| < eps_width
+                    # Shifted tanh outside the deadzone
+                    # Use smooth approximation to avoid discontinuity:
+                    # sign_factor = -tanh(steep * (|dev| - eps) * sign(dev) / A)
+                    #             * smoothstep(|dev|, eps)
+                    abs_dev = torch.abs(deviation)
+                    shifted_dev = abs_dev - eps_width
+                    shifted_dev_signed = shifted_dev * torch.sign(deviation)
+
+                    # Smooth transition at deadzone boundary using sigmoid
+                    # transition_width controls sharpness of deadzone edge
+                    transition_width = 0.1 * A_ref  # smooth over 10% of A
+                    gate = torch.sigmoid((abs_dev - eps_width) / (transition_width + 1e-8))
+
+                    sign_factor = -torch.tanh(steepness * shifted_dev_signed / (A_ref + 1e-6)) * gate
+                else:
+                    # Standard CTC (no deadzone)
+                    sign_factor = -torch.tanh(steepness * deviation / (A_ref + 1e-6))
+
+                velocity_raw = velocity_raw * sign_factor
+
+            # 3. Velocity-dependent fp drag (Tranquillo & Lauffenburger 1987)
+            if self.fp_drag > 0:
+                vel_i = x_i[:, 1+self.dimension:1+2*self.dimension]
+                speed = torch.sqrt(torch.sum(vel_i**2, dim=1, keepdim=True))
+                drag_factor = 1.0 / (1.0 + self.fp_drag * speed / self.v_ref)
+                velocity_raw = velocity_raw * drag_factor
+
+            return velocity_raw
+
+        elif mode == 'pf':
+            weights = torch.exp(-dist**2 / (2 * (self.influence_radius/3)**2))
+
+            if parameters_i is not None:
+                consumption = parameters_i[:, 2]
+                production = parameters_i[:, 3]
+            else:
+                consumption = self.consumption_rate
+                production = self.production_rate
+
+            field_updates = torch.zeros((pos_i.size(0), 2), device=pos_i.device)
+            field_updates[:, 0] = -consumption * weights
+            field_updates[:, 1] = production * weights
+            return field_updates
+
+        else:  # mode == 'pp'
+            if parameters_i is not None:
+                p1 = parameters_i[:, 4]
+                p2 = parameters_i[:, 5]
+                p3 = parameters_i[:, 6]
+                p4 = parameters_i[:, 7]
+
+                f = (p1 * torch.exp(-dist ** (2 * p2) / (2 * self.sigma ** 2))
+                     - p3 * torch.exp(-dist ** (2 * p4) / (2 * self.sigma ** 2)))
+
+                forces = f[:, None] * d_pos / dist_safe.unsqueeze(1)
+            else:
+                forces = torch.zeros_like(pos_i)
+                in_range = dist < self.repulsion_range
+                if in_range.any():
+                    dir_norm = d_pos / dist_safe.unsqueeze(1)
+                    repulsion_mag = self.repulsion_strength * torch.exp(
+                        -5.0 * dist[in_range] / self.repulsion_range
+                    )
+                    forces[in_range] = -dir_norm[in_range] * repulsion_mag.unsqueeze(1)
+
+            # Field-dependent pp damping (Painter & Hillen 2002)
+            if self.pp_damping > 0 and self.ctc_threshold > 0:
+                C1_local = x_i[:, 6:7].squeeze(1)
+                A_ref = self.A_ref
+                base_T = self.ctc_threshold * A_ref
+
+                if (parameters_i is not None and self.cross_type_factor > 0
+                        and x_i.numel() > 0):
+                    type_i = x_i[:, 1 + 2*self.dimension].long()
+                    n_types = type_i.max().item() + 1 if type_i.numel() > 0 else 1
+                    mean_idx = (n_types - 1) / 2.0
+                    type_offset = self.cross_type_factor * (type_i.float() - mean_idx)
+                    T_local = base_T * (1.0 + type_offset)
+                else:
+                    T_local = base_T
+
+                width = self.pp_damping_width * A_ref
+                deviation = (C1_local - T_local)
+                damping_factor = 1.0 - self.pp_damping * torch.exp(-deviation**2 / (2 * width**2 + 1e-8))
+                forces = forces * damping_factor.unsqueeze(1)
+
+            return forces
+
+    def update(self, aggr_out, mode=None):
+        if mode == 'interpolate':
+            C1_weighted = aggr_out[:, 0:1]
+            C2_weighted = aggr_out[:, 1:2]
+            weight_sum = aggr_out[:, 2:3]
+            weight_sum = torch.clamp(weight_sum, min=1e-10)
+            return torch.cat([C1_weighted / weight_sum, C2_weighted / weight_sum], dim=1)
+        else:
+            return aggr_out