"""Material-conditioned DENIM constitutive core.

The module retains the small-strain elastic/J2/associative-flow return map
used by DENIM and conditions only the unknown hardening closure on a compact
material descriptor.  It deliberately does not encode the closed-form J2 or
Chaboche hardening equations.
"""

from __future__ import annotations

from dataclasses import dataclass

import torch
from torch import nn
from torch.nn import functional as F

from src.t2_graybox_discrete_energy import (
    GrayboxState,
    deviatoric,
    double_contract,
    elastic_stress,
    initial_state,
    von_mises,
)


DESCRIPTOR_NAMES = (
    "young_scaled",
    "poisson",
    "yield_scaled",
    "linear_isotropic_scaled",
    "backstress_c1_scaled",
    "backstress_gamma1_scaled",
    "backstress_c2_scaled",
    "backstress_gamma2_scaled",
    "isotropic_saturation_scaled",
    "isotropic_rate_scaled",
    "family_j2",
    "family_chaboche",
    "family_incomplete",
)


def material_descriptor(parameters: dict[str, float | str]) -> torch.Tensor:
    """Build a dimensionless descriptor without exposing hidden DENIM laws."""

    family = str(parameters["material_model"])
    indicators = {
        "j2_linear_isotropic": (1.0, 0.0, 0.0),
        "chaboche_combined": (0.0, 1.0, 0.0),
        "hidden_three_memory_tabulated_hardening": (0.0, 0.0, 1.0),
    }
    if family not in indicators:
        raise ValueError(f"Unsupported material family: {family}")
    # The incomplete family exposes only E, nu and initial yield stress.  The
    # zero entries are intentional, not missing-data imputation.
    hidden = family == "hidden_three_memory_tabulated_hardening"
    value = (
        float(parameters["young_pa"]) / 200.0e9,
        float(parameters["poisson"]),
        float(parameters["yield_stress_pa"]) / 300.0e6,
        0.0 if hidden else float(parameters.get("hardening_modulus_pa", 0.0)) / 5.0e9,
        0.0 if hidden else float(parameters.get("backstress_c1_pa", 0.0)) / 30.0e9,
        0.0 if hidden else float(parameters.get("backstress_gamma1", 0.0)) / 80.0,
        0.0 if hidden else float(parameters.get("backstress_c2_pa", 0.0)) / 10.0e9,
        0.0 if hidden else float(parameters.get("backstress_gamma2", 0.0)) / 20.0,
        0.0 if hidden else float(parameters.get("isotropic_saturation_pa", 0.0)) / 100.0e6,
        0.0 if hidden else float(parameters.get("isotropic_rate", 0.0)) / 10.0,
        *indicators[family],
    )
    return torch.tensor(value, dtype=torch.float32)


class MaterialEncoder(nn.Module):
    def __init__(self, descriptor_size: int = len(DESCRIPTOR_NAMES), hidden: int = 32):
        super().__init__()
        self.network = nn.Sequential(
            nn.Linear(descriptor_size, hidden),
            nn.SiLU(),
            nn.Linear(hidden, hidden),
            nn.SiLU(),
        )

    def forward(self, descriptor: torch.Tensor) -> torch.Tensor:
        return self.network(descriptor)


class ConditionalIsotropicHardening(nn.Module):
    """Descriptor-conditioned, zero-anchored monotone hardening curve."""

    def __init__(self, embedding: int = 32, neurons: int = 12, plastic_scale: float = 0.01):
        super().__init__()
        self.neurons = int(neurons)
        self.plastic_scale = float(plastic_scale)
        self.parameter_head = nn.Linear(embedding, 2 * neurons + 1)

    def _positive_parameters(
        self, embedding: torch.Tensor
    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        raw = self.parameter_head(embedding)
        weight = 0.45 * torch.sigmoid(raw[..., : self.neurons])
        slope = F.softplus(raw[..., self.neurons : 2 * self.neurons]) + 1.0e-5
        linear = 0.60 * torch.sigmoid(raw[..., -1])
        return weight, slope, linear

    def forward(
        self,
        peeq: torch.Tensor,
        stress_scale: torch.Tensor,
        embedding: torch.Tensor,
    ) -> torch.Tensor:
        weight, slope, linear = self._positive_parameters(embedding)
        coordinate = peeq[..., None] / self.plastic_scale
        saturation = -torch.expm1(-slope * coordinate)
        dimensionless = linear * coordinate.squeeze(-1) + (weight * saturation).sum(dim=-1)
        return stress_scale * dimensionless


class ConditionalDENIM(nn.Module):
    """Shared neural hardening closure conditioned on material metadata."""

    def __init__(self, channels: int = 2, embedding: int = 32, hidden: int = 48):
        super().__init__()
        self.channels = int(channels)
        self.encoder = MaterialEncoder(hidden=embedding)
        self.isotropic = ConditionalIsotropicHardening(embedding=embedding)
        self.modulus_head = nn.Sequential(
            nn.Linear(embedding, hidden), nn.SiLU(), nn.Linear(hidden, channels)
        )
        state_size = 3 + 3 * channels + embedding
        self.recovery = nn.Sequential(
            nn.Linear(state_size, hidden),
            nn.SiLU(),
            nn.Linear(hidden, hidden),
            nn.SiLU(),
            nn.Linear(hidden, channels),
        )

    def encode(self, descriptor: torch.Tensor) -> torch.Tensor:
        return self.encoder(descriptor)

    def moduli(self, stress_scale: torch.Tensor, embedding: torch.Tensor) -> torch.Tensor:
        # 0--400 times the initial yield stress covers the synthetic families
        # while retaining a bounded, nonnegative kinematic modulus.
        return 400.0 * stress_scale[..., None] * torch.sigmoid(self.modulus_head(embedding))

    def isotropic_radius(
        self,
        peeq: torch.Tensor,
        stress_scale: torch.Tensor,
        embedding: torch.Tensor,
    ) -> torch.Tensor:
        return self.isotropic(peeq, stress_scale, embedding)

    def update_memories(
        self,
        state: GrayboxState,
        flow_direction: torch.Tensor,
        increment: torch.Tensor,
        stress_scale: torch.Tensor,
        embedding: torch.Tensor,
    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        radius = self.isotropic_radius(state.peeq, stress_scale, embedding)
        old_norm = torch.sqrt(torch.clamp(double_contract(state.previous_flow, state.previous_flow), min=0.0))
        current_norm = torch.sqrt(torch.clamp(double_contract(flow_direction, flow_direction), min=0.0))
        reversal = double_contract(state.previous_flow, flow_direction) / (
            old_norm * current_norm
        ).clamp_min(1.0e-12)
        scale = stress_scale.clamp_min(1.0)
        norms = torch.sqrt(torch.clamp(double_contract(state.memories, state.memories), min=0.0)) / scale[..., None]
        projections = double_contract(state.memories, flow_direction[..., None, :]) / scale[..., None]
        cross = torch.zeros_like(norms)
        if self.channels > 1:
            for index in range(self.channels):
                other = (index + 1) % self.channels
                cross[..., index] = double_contract(
                    state.memories[..., index, :], state.memories[..., other, :]
                ) / scale.square()
        scalars = torch.stack((state.peeq / 0.05, radius / scale, reversal), dim=-1)
        features = torch.cat((scalars, norms, projections, cross, embedding), dim=-1)
        recovery = 250.0 * torch.sigmoid(self.recovery(features)) + 1.0e-7
        moduli = self.moduli(stress_scale, embedding)
        numerator = state.memories + (
            (2.0 / 3.0)
            * moduli[..., :, None]
            * increment[..., None, None]
            * flow_direction[..., None, :]
        )
        denominator = 1.0 + recovery * increment[..., None]
        return numerator / denominator[..., None], recovery, moduli


def _candidate_update(
    trial_deviatoric: torch.Tensor,
    state: GrayboxState,
    increment: torch.Tensor,
    shear: torch.Tensor,
    yield_stress: torch.Tensor,
    embedding: torch.Tensor,
    law: ConditionalDENIM,
    *,
    direction_iterations: int = 6,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
    shifted = trial_deviatoric - state.backstress
    direction = 1.5 * shifted / von_mises(shifted).clamp_min(1.0)[..., None]
    memories = state.memories
    recovery = moduli = None
    for _ in range(direction_iterations):
        memories, recovery, moduli = law.update_memories(
            state, direction, increment, yield_stress, embedding
        )
        denominator = 1.0 + recovery * increment[..., None]
        effective = trial_deviatoric - (
            state.memories / denominator[..., None]
        ).sum(dim=-2)
        direction = 1.5 * effective / von_mises(effective).clamp_min(1.0)[..., None]
    memories, recovery, moduli = law.update_memories(
        state, direction, increment, yield_stress, embedding
    )
    denominator = 1.0 + recovery * increment[..., None]
    effective = trial_deviatoric - (state.memories / denominator[..., None]).sum(dim=-2)
    radius = yield_stress + law.isotropic_radius(
        state.peeq + increment, yield_stress, embedding
    )
    effective_modulus = 3.0 * shear + (moduli / denominator).sum(dim=-1)
    residual = von_mises(effective) - effective_modulus * increment - radius
    return residual, direction, memories, recovery, moduli


def advance(
    strain: torch.Tensor,
    state: GrayboxState,
    young: torch.Tensor,
    poisson: torch.Tensor,
    yield_stress: torch.Tensor,
    descriptor: torch.Tensor,
    law: ConditionalDENIM,
    *,
    bisection_iterations: int = 24,
    direction_iterations: int = 6,
) -> tuple[torch.Tensor, GrayboxState, dict[str, torch.Tensor]]:
    embedding = law.encode(descriptor)
    trial = elastic_stress(strain, state.plastic_strain, young, poisson)
    trial_deviatoric = deviatoric(trial)
    shifted = trial_deviatoric - state.backstress
    old_radius = yield_stress + law.isotropic_radius(state.peeq, yield_stress, embedding)
    trial_function = von_mises(shifted) - old_radius
    plastic = trial_function > yield_stress.clamp_min(1.0) * 1.0e-12
    shear = young / (2.0 * (1.0 + poisson))
    lower = torch.zeros_like(trial_function)
    upper = 2.0 * F.relu(trial_function) / (3.0 * shear).clamp_min(1.0) + 1.0e-14
    for _ in range(16):
        residual, *_ = _candidate_update(
            trial_deviatoric, state, upper, shear, yield_stress, embedding, law,
            direction_iterations=direction_iterations,
        )
        upper = torch.where(plastic & (residual > 0.0), 2.0 * upper, upper)
    for _ in range(bisection_iterations):
        middle = 0.5 * (lower + upper)
        residual, *_ = _candidate_update(
            trial_deviatoric, state, middle, shear, yield_stress, embedding, law,
            direction_iterations=direction_iterations,
        )
        lower = torch.where(plastic & (residual > 0.0), middle, lower)
        upper = torch.where(plastic & (residual <= 0.0), middle, upper)
    increment = torch.where(plastic, 0.5 * (lower + upper), torch.zeros_like(lower))
    residual, direction, memories, recovery, moduli = _candidate_update(
        trial_deviatoric, state, increment, shear, yield_stress, embedding, law,
        direction_iterations=direction_iterations,
    )
    direction = torch.where(plastic[..., None], direction, torch.zeros_like(direction))
    memories = torch.where(plastic[..., None, None], memories, state.memories)
    updated = GrayboxState(
        plastic_strain=state.plastic_strain + increment[..., None] * direction,
        peeq=state.peeq + increment,
        memories=memories,
        previous_flow=torch.where(plastic[..., None], direction, state.previous_flow),
    )
    stress = elastic_stress(strain, updated.plastic_strain, young, poisson)
    return stress, updated, {
        "plastic_increment": increment,
        "yield_residual": torch.where(plastic, residual, torch.zeros_like(residual)),
        "recovery": recovery,
        "moduli": moduli,
        "plastic": plastic,
    }


def rollout(
    strain: torch.Tensor,
    young: torch.Tensor,
    poisson: torch.Tensor,
    yield_stress: torch.Tensor,
    descriptor: torch.Tensor,
    law: ConditionalDENIM,
    *,
    bisection_iterations: int = 24,
) -> dict[str, torch.Tensor]:
    batch, points, _ = strain.shape
    state = initial_state(batch, channels=law.channels, dtype=strain.dtype, device=strain.device)
    result: dict[str, list[torch.Tensor]] = {
        name: []
        for name in (
            "stress", "plastic_strain", "peeq", "memories",
            "plastic_increment", "yield_residual", "recovery", "moduli",
        )
    }
    for point in range(points):
        stress, state, diagnostics = advance(
            strain[:, point], state, young, poisson, yield_stress, descriptor, law,
            bisection_iterations=bisection_iterations,
        )
        result["stress"].append(stress)
        result["plastic_strain"].append(state.plastic_strain)
        result["peeq"].append(state.peeq)
        result["memories"].append(state.memories)
        for name in ("plastic_increment", "yield_residual", "recovery", "moduli"):
            result[name].append(diagnostics[name])
    return {name: torch.stack(values, dim=1) for name, values in result.items()}


__all__ = [
    "ConditionalDENIM",
    "DESCRIPTOR_NAMES",
    "advance",
    "material_descriptor",
    "rollout",
]