How shengfeng233/pinn-preprocess works
Project type and purpose
• This is a scientific preprocessing library and demonstration code for Physics-Informed Neural Networks (PINNs). It provides routines to non-dimensionalize PDE data, build and train PINN models with different normalization schemes, validate results, and plot training analyses.
File structure
• pinn_model.py: defines network architectures and normalization variants (Baseline, InnerNormalization, Normalization).
• train_sweep.py: orchestrates parameter sweeps and model training runs.
• utilities.py: helper functions for data loading, transformation, and normalization.
• validation_func.py: routines to compute error metrics and compare model outputs against reference data.
• plot_bar_batchsize_ave.py: generates bar-charts of average metrics versus batch size.
• README.md & LICENSE: project overview and license.
• No subdirectories—each script represents a major module in the pipeline.
Extract key insights from README
• Three normalization modes are supported.
• Input/output data must be downloaded externally.
• Focus is on improving training by data transformation.
Components
a. Main components
• External Data Source
– Raw PDE data (training + reference) stored on Google Drive
• Data Ingestion & Preprocessing
– utilities.py: loads data, applies linear scaling
– Three transformation modes (Baseline, InnerNormalization, Normalization)
• PINN Model Definition
– pinn_model.py: defines network graph, loss terms (PDE residual, data mismatch)
• Training Orchestrator
– train_sweep.py: sets hyperparameters, loops over experiments, launches training
• Validation Module
– validation_func.py: computes L2 errors, residual norms
• Visualization Module
– plot_bar_batchsize_ave.py: reads validation outputs, plots performance vs batch size
b. Relationships & interactions
1. External Data → Data Ingestion & Preprocessing
2. Preprocessed Data → PINN Model Definition
3. Model + Hyperparams → Training Orchestrator → produces Model Checkpoints & Logs
4. Checkpoints + Reference Data → Validation Module → Validation Metrics
5. Validation Metrics → Visualization Module → Plots
c. Architectural patterns & design principles
• Modular pipeline: each script is a single responsibility.
• Data-flow orientation: clear stages from data ingestion → transformation → model → evaluation → plotting.
• Extensibility: separate normalization modes for easy extension to other PDEs.
d. Key technologies & libraries
• Python scientific stack (NumPy, possibly TensorFlow/PyTorch inside pinn_model.py)
• Command-line orchestration (train_sweep.py)
• Matplotlib or similar for plots