MultiFidelityTrainingHistory<T>

Models & Types ยท AiDotNet.PhysicsInformed

Training history for multi-fidelity PINN training.

Inheritance: TrainingHistory<T> โ†’ MultiFidelityTrainingHistory<T>

Implements: IMultiFidelityTrainingHistory<T>

How It Works

For Beginners: This class tracks the training progress of a multi-fidelity PINN. It extends the base TrainingHistory with additional metrics specific to multi-fidelity learning:

  • LowFidelityLosses: Error on cheap/approximate data
  • HighFidelityLosses: Error on expensive/accurate data
  • CorrelationLosses: How well the model captures the relationship between fidelity levels
  • PhysicsLosses: PDE residual errors

Typical Training Dynamics:

  1. Early training: Low-fidelity loss dominates (most data)
  2. Mid training: High-fidelity becomes important (precision matters)
  3. Late training: Physics loss should be low (PDE satisfied)

Constructors

Constructor Summary
MultiFidelityTrainingHistory()

Properties

Property Summary
List<T> CorrelationLosses { get; } Gets the correlation losses per epoch.
List<T> HighFidelityLosses { get; } Gets the high-fidelity data losses per epoch.
List<T> LowFidelityLosses { get; } Gets the low-fidelity data losses per epoch.
List<T> PhysicsLosses { get; } Gets the PDE residual losses per epoch.

Methods

Method Summary
void AddEpoch(T totalLoss, T lowFidelityLoss, T highFidelityLoss, T correlationLoss, T physicsLoss) Records metrics for a training epoch.