MultiScalePINN<T>

Models & Types · AiDotNet.PhysicsInformed.PINNs

Implements a Multi-Scale Physics-Informed Neural Network for solving PDEs with multiple scales.

Inheritance: NeuralNetworkBase<T> → MultiScalePINN<T>

How It Works

For Beginners: Multi-scale problems are challenging because features span vastly different sizes. A single neural network struggles to capture both large-scale trends and fine details.

Solution: Multi-Scale PINN Uses multiple sub-networks, each specialized for a different scale:

  • Coarse network: Captures large-scale, smooth variations
  • Fine network(s): Capture small-scale details and fluctuations

Architecture: Input (x,t) → [Coarse Net] → u_coarse(x,t) → [Fine Net 1] → u_fine1(x,t) → [Fine Net 2] → u_fine2(x,t) → …

Total solution: u(x,t) = u_coarse + u_fine1 + u_fine2 + …

Training Strategy:

  1. Progressive Training: Train coarse first, then add finer scales
  2. Simultaneous Training: Train all scales together with adaptive weights
  3. Alternating Training: Alternate between scales during training

Key Features:

  • Fourier feature encoding at different frequencies for each scale
  • Adaptive loss weighting to balance scale contributions
  • Scale coupling terms to ensure consistency
  • Progressive activation of finer scales during training

Applications:

  • Turbulence modeling (large eddies + small vortices)
  • Composite materials (macroscopic + fiber-scale behavior)
  • Multi-physics problems (thermal + mechanical + chemical)
  • Climate modeling (global + regional + local scales)

Example

var architecture = new NeuralNetworkArchitecture<double>(
    InputType.OneDimensional, NeuralNetworkTaskType.Regression,
    inputSize: 2, outputSize: 1);
            
// The same physics at a coarse and a fine length scale, coupled where they overlap.
var pde = new MultiScaleReactionDiffusion<double>(
    coarseDiffusivity: 1.0,
    fineDiffusivity: 0.01);
            
// Held at zero on both edges of the domain.
var boundaryConditions = new IBoundaryCondition<double>[]
{
    new DirichletBoundaryCondition<double>(boundaryValue: 0.0, lowerBound: 0.0, upperBound: 1.0)
};
            
var msPinn = new MultiScalePINN<double>(architecture, pde, boundaryConditions);

Constructors

Constructor Summary
MultiScalePINN(NeuralNetworkArchitecture<T> architecture, IMultiScalePDE<T> multiScalePDE, IBoundaryCondition<T>[] boundaryConditions, IInitialCondition<T> initialCondition, MultiScaleTrainingOptions<T> trainingOptions, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer, MultiScalePINNOptions options) Initializes a new instance of the Multi-Scale PINN.

Properties

Property Summary
T[] ScaleCharacteristicLengths { get; } Gets the characteristic length scales.
bool SupportsTraining { get; }
int NumberOfScales { get; } Gets the number of scales in this multi-scale PINN.

Methods

Method Summary
ModelMetadata<T> GetModelMetadata()
Tensor<T> Forward(Tensor<T> input) Forward pass through all scale networks: the sum of the active scales’ outputs.
Tensor<T> ForwardForTraining(Tensor<T> input)
Tensor<T> ForwardWithMemory(Tensor<T> input)
TrainingHistory<T> Solve(int epochs, double learningRate, bool verbose) Solves the multi-scale PDE using physics-informed training.
Vector<T> GetGradients()