examples/polymer_optimization.py

"""
Example: Physics-Informed Bayesian Optimization for Polymer Design

This example demonstrates optimizing a polymer formulation where:
- A physics model (simplified Flory-Huggins + Arrhenius kinetics) provides
  prior knowledge about how composition and temperature affect properties.
- Initial experimental data provides a warm start.
- The BO loop efficiently explores the design space, leveraging both
  physics and data to minimize the number of experiments needed.

Objective: Maximize polymer recyclability metric (higher is better).
Parameters:
  - monomer_ratio: Ratio of monomer A to B (0.1 to 0.9)
  - temperature: Reaction temperature in Kelvin (350 to 500)
  - catalyst_loading: Catalyst weight percent (0.5 to 5.0)
"""

import torch
from torch import Tensor

from physics_informed_bo.experiment.parameter_space import ParameterSpace
from physics_informed_bo.experiment.campaign import OptimizationCampaign
from physics_informed_bo.config import OptimizationConfig, AcquisitionType


# ============================================================================
# 1. Define the physics model (simplified polymer recyclability model)
# ============================================================================

def polymer_physics_model(X: Tensor) -> Tensor:
    """Simplified physics model for polymer recyclability.

    Based on:
    - Flory-Huggins mixing thermodynamics
    - Arrhenius reaction kinetics
    - Empirical catalyst efficiency model

    Args:
        X: Tensor of shape (n, 3) with columns:
           [monomer_ratio, temperature, catalyst_loading]

    Returns:
        Predicted recyclability metric (higher = better).
    """
    ratio = X[:, 0]       # monomer ratio
    temp = X[:, 1]         # temperature (K)
    catalyst = X[:, 2]     # catalyst loading (wt%)

    # Flory-Huggins: optimal mixing near 50:50 ratio
    chi = 0.5 - 0.3 * (ratio - 0.5) ** 2  # chi parameter
    mixing_term = -ratio * torch.log(ratio + 1e-8) - (1 - ratio) * torch.log(1 - ratio + 1e-8)
    mixing_free_energy = mixing_term - chi * ratio * (1 - ratio)

    # Arrhenius: reaction rate dependence on temperature
    Ea = 50.0  # kJ/mol activation energy
    R = 8.314e-3  # kJ/(mol·K)
    rate = torch.exp(-Ea / (R * temp))

    # Catalyst efficiency (diminishing returns)
    catalyst_eff = 1 - torch.exp(-0.8 * catalyst)

    # Combined recyclability metric
    recyclability = 5.0 * mixing_free_energy * rate * catalyst_eff + 2.0

    return recyclability


# ============================================================================
# 2. Define the "true" function (simulates real experiments with noise)
# ============================================================================

def true_recyclability(params: dict) -> float:
    """Simulate running an actual experiment (physics + discrepancy + noise)."""
    X = torch.tensor(
        [[params["monomer_ratio"], params["temperature"], params["catalyst_loading"]]],
        dtype=torch.float64,
    )

    # Physics prediction
    physics = polymer_physics_model(X).item()

    # Add model discrepancy (physics doesn't capture everything)
    ratio = params["monomer_ratio"]
    temp = params["temperature"]
    discrepancy = 0.3 * torch.sin(torch.tensor(10.0 * ratio)).item() \
                  + 0.1 * (temp - 400) / 100

    # Add measurement noise
    noise = 0.05 * torch.randn(1).item()

    return physics + discrepancy + noise


# ============================================================================
# 3. Set up the optimization campaign
# ============================================================================

def main():
    # Define parameter space
    space = ParameterSpace()
    space.add_continuous("monomer_ratio", 0.1, 0.9, units="ratio")
    space.add_continuous("temperature", 350.0, 500.0, units="K")
    space.add_continuous("catalyst_loading", 0.5, 5.0, units="wt%")

    # Generate some initial experimental data
    torch.manual_seed(42)
    X_init = space.sample_latin_hypercube(5)
    y_init = torch.tensor(
        [true_recyclability(space.to_dict(X_init)[i]) for i in range(5)],
        dtype=torch.float64,
    ).unsqueeze(-1)

    print("=== Initial Data ===")
    for i, (params, y_val) in enumerate(zip(space.to_dict(X_init), y_init)):
        print(f"  Exp {i+1}: {params} -> {y_val.item():.4f}")

    # Configure optimization
    config = OptimizationConfig(
        acquisition_type=AcquisitionType.PHYSICS_INFORMED_EI,
        n_initial_samples=5,
        max_iterations=20,
        use_physics_mean=True,
        noise_variance=0.01,
    )

    # Create campaign
    campaign = OptimizationCampaign(
        name="polymer_recyclability",
        parameter_space=space,
        physics_fn=polymer_physics_model,
        initial_data=(X_init, y_init),
        config=config,
        maximize=True,
    )

    print("\n=== Running Optimization ===")

    def callback(iteration, best):
        print(f"  Iteration {iteration}: best = {best['objective']:.4f}")

    # Run automated optimization
    results_df = campaign.run_automated(
        objective_fn=true_recyclability,
        max_iterations=15,
        callback=callback,
    )

    # Report results
    best = campaign.get_best()
    print(f"\n=== Best Result ===")
    print(f"  Parameters: {best['parameters']}")
    print(f"  Objective:  {best['objective']:.4f}")

    print(f"\n=== Campaign Summary ===")
    summary = campaign.summary()
    print(f"  Total experiments: {summary['n_experiments']}")
    print(f"  Physics model R²: {summary['model_summary'].get('model_quality', {}).get('r2', 'N/A')}")

    # Save results
    campaign.save("polymer_campaign.json")
    results_df.to_csv("polymer_results.csv", index=False)
    print("\nResults saved to polymer_campaign.json and polymer_results.csv")


if __name__ == "__main__":
    main()