Model Fitting

This module uses the JAX library for high-performance, gradient-based optimization to fit the parameters of a FosterNetwork to thermal impedance data.

fit_foster_network()

fit_foster_network(
  time_data,
  impedance_data,
  n_layers,
  config=None,
  initialization_type='exponential',
  random_seed=0
)

Fits an N-layer Foster network to thermal impedance data.

• time_data: array-like - Array of time points.
• impedance_data: array-like - Array of corresponding thermal impedance values.
• n_layers: int - The number of RC layers to fit.
• config: OptimizationConfig, optional - Configuration for the optimization process (e.g., optimizer, learning rate, tolerances). If None, default settings are used.
• initialization_type: str - Method for generating the initial guess (‘uniform’ or ‘exponential’).
• random_seed: int - Seed for randomizing the initial guess.
• Returns: FosterModelResult - An object containing the fitted FosterNetwork and metadata about the optimization.

fit_optimal_foster_network()

fit_optimal_foster_network(
  time_data,
  impedance_data,
  max_layers=5,
  selection_criterion='bic',
  config=None,
  initialization_type='exponential',
  random_seed=0
)

Fits models with 1 to max_layers and selects the best one using a model selection criterion (AIC or BIC). This is useful for automatically determining the required complexity of the model.

• time_data: array-like - Array of time points.
• impedance_data: array-like - Array of corresponding thermal impedance values.
• max_layers: int - The maximum number of layers to test.
• selection_criterion: str - 'aic' or 'bic'. The criterion used for model selection.
• config: OptimizationConfig, optional - Configuration for the optimization process.
• initialization_type: str - Method for generating the initial guess (‘uniform’ or ‘exponential’).
• random_seed: int - Seed for randomizing the initial guess.
• Returns: EvaluatedFosterModelResult - The best model found, including its selection criteria scores.

OptimizationConfig

A data class for configuring the fitting process.

• optimizer: str - 'lbfgs' or 'adam'.
• n_steps: int - Maximum number of optimization steps.
• learning_rate: float - Learning rate for the Adam optimizer.
• loss_tol: float- Convergence tolerance for the change in loss value.
• gradient_tol: float - Convergence tolerance for the gradient norm.
• params_rtol: float - Relative tolerance for parameter change.
• params_atol: float - Absolute tolerance for parameter change.
• randomize_guess_strength: float - Stddev of multiplicative noise for randomizing the initial guess. 0 implies a deterministic initial guess.

Example: Automatic Model Selection

1. Generate Synthetic Data

   First, we create a “true” 3-layer network and generate noisy impedance data from it, simulating a real-world measurement.

   import numpy as np
   from thermal_network.networks import FosterNetwork
   from thermal_network.impedance import foster_impedance_time_domain
   
   # Define a true 3-layer network
   true_network = FosterNetwork(r_values=[0.2, 0.8, 0.5], c_values=[15.0, 1.0, 4.0])
   time_vec = np.logspace(-1, 2, 200)
   true_impedance = foster_impedance_time_domain(true_network, time_vec)
   
   # Add noise
   np.random.seed(42)
   noise_level = 0.015 * np.max(true_impedance)
   noisy_impedance = true_impedance + noise_level * np.random.randn(true_impedance.shape[0])

2. Find the Optimal Model

   We use fit_optimal_foster_network to test models from 1 to 5 layers and let it choose the best one based on the Bayesian Information Criterion (BIC).

   from thermal_network.fitting import fit_optimal_foster_network
   
   # Find the optimal model up to 5 layers
   optimal_model = fit_optimal_foster_network(
       time_vec,
       noisy_impedance,
       max_layers=5,
       selection_criterion='bic',
   )
   
   print(f"Optimal model found with {optimal_model.n_layers} layers.")
   print(f"BIC Score: {optimal_model.selection_criteria['bic']:.2f}")
   print("Fitted Network:", optimal_model.network)