How nguyenkhoa0209/pinns_tutorial works
Project type and purpose
• Project type: An educational/open-source tutorial library for Physics‐Informed Neural Networks (PINNs).
• Purpose: Provide code modules, Jupyter notebooks, and data to demonstrate how to solve PDEs via data‐driven ML and PINNs, including advanced sampling strategies.
File structure
• Top‐level modules:
• data/ (raw numerical data for training & validation)
• networks.py (neural network architectures)
• pinns.py & pinns_draft_lbfgs.py (PINN training engines, loss definitions)
• adapt_sampling.py (adaptive or residual‐based collocation‐point selection)
• utils.py (data loading, normalization, visualization helpers)
• *.ipynb notebooks (tutorial entry points: part1, part2, solutions)
• slides/ (PDF and text docs for background theory)
• Patterns: Modular Python scripts for core functionality; notebooks orchestrate workflows.
README for insights
• Two phases: Part 1 (pure data‐driven ML) and Part 2 (PINNs for PDEs).
• Online execution via Colab/Binder—emphasizes notebooks as front end.
• References indicate focus on sampling strategies, training stability, advanced PINN features.
Main components & interactions
a. Components:
1. Data Store (data/ folder)
2. Preprocessing & Utilities (utils.py)
3. Network Definition (networks.py)
4. PINN Engine (pinns.py & pinns_draft_lbfgs.py)
5. Sampling Module (adapt_sampling.py)
6. Notebook Interface (part1_*.ipynb, part2_*.ipynb, solution_*.ipynb)
7. Slide Decks (slides/)
b. Relationships:
• Notebooks import utils, networks, pinns, and adapt_sampling
• utils loads data from data/ → preprocess → feeds inputs to PINN Engine
• networks defines f_θ(x) → PINN Engine applies physics‐informed losses
• adapt_sampling provides collocation‐points to PINN Engine
c. Architectural patterns:
• Layered modular design (data → utils → core modules → interface)
• Separation of concerns: network architecture, solver logic, sampling strategy
• Extensibility points: swap network architectures, sampling methods, optimizers
d. Key technologies:
• Python 3.x, NumPy, TensorFlow or PyTorch (in engines)
• Jupyter notebooks as primary UIs
• Adaptive sampling algorithms for PINNs
Detail but avoid over‐fragmentation
• Every Python script as a module block
• Import/use relationships rather than internal function calls
• That notebooks are the entry points orchestrating workflows
[Data Store] → [Utilities Module] →
↘
[PINN Engine] ← [Network Definition]
↑
[Sampling Module]
[Notebooks] ⟲ imports {Utilities, Network, PINN Engine, Sampling} → visualizes outputs
[Slides] (external docs)