How vanthienphan2004/physics-informed-neural-networks-for-collective-cancer-invasion works
Project Type and Purpose
• This is a research‐oriented Python application implementing physics‐informed neural networks (PINNs) and Fourier neural operators (FNOs) to solve coupled reaction‐diffusion PDEs modeling collective cancer invasion.
• It is not a web or distributed microservice but a sequential simulation pipeline: configuration → data generation → model construction → training (with PDE constraints) → results storage → visualization.
File Structure Highlights
• Entry point: main.py
• Orchestrator: controller.py (implements Singleton SimulationController)
• Configuration: config.json, loaded at startup
• Scripts directory (scripts/):
• data_loader.py (synthetic data + IC/BC enforcement)
• models.py (PINN & FNO architectures)
• trainer.py (physics‐informed loss, PDE residuals, optimization)
• visualization.py (plotting, saving snapshots)
• utils.py (device detection, common helpers)
• Output directories: results/ (weights, predictions, summaries), plots/ (visualizations)
Core Components
a. Configuration Manager
• Loads config.json; exposes parameters (model type, PDE constants, training hyperparameters).
b. SimulationController (Singleton)
• Central orchestrator; instantiated by main.py
• Sequence: load config → initialize data pipeline → initialize model → initialize trainer → run training → invoke visualization → save outputs
c. Data Pipeline
• data_loader.py generates initial/boundary condition data, prepares PyTorch tensors
d. Model Architectures
• models.py defines two types: PINN and FNO classes (e.g., CoupledSpeciesTFNO)
e. Training Engine
• trainer.py computes physics‐informed losses, backpropagation, optimizer steps, monitors metrics
f. Results Storage
• File system writes: model_weights.pth, predictions.{npz,csv}, results_summary.json
g. Visualization Module
• visualization.py reads predictions → generates plots → writes to plots/
Relationships and Data Flow
• main.py → SimulationController
• SimulationController
↓ loads
Configuration Manager
↓ passes parameters to
┌───────────────┬───────────────┐
│ Data Pipeline │ Model Factory │
└──────┬────────┴──────┬────────┘
↓ ↓
Training Engine ←───┘
↓
Results Storage
↓
Visualization Module
↓
plots/
Architectural Patterns and Principles
• Singleton Pattern: SimulationController ensures single orchestrator instance
• Modular Design: separation of data, model, trainer, visualization
• Configuration-Driven: all parameters in JSON, injected at runtime
• Pipeline Pattern: sequential processing stages
• Type Safety & Exception Handling: clear module boundaries
Technologies and Frameworks
• Python 3.12+
• PyTorch 2.9+ (with optional CUDA)
• NumPy for data arrays
• Matplotlib / seaborn for plotting
• JSON for configuration and results summaries