How ehsanhaghighat/en-deeponet works
Project Type & Purpose
• This is a machine-learning research repository implementing and evaluating a novel neural-operator architecture (“En-DeepONet”) across several PDE/test-case experiments (Laplace, Level-set/Eikonal, Marmousi, OpenFWI, SmoothVelocity).
• The README confirms the goal: train/enhance DeepONet to learn moving-solution operators for seismic applications and compare performance.
File Structure
• Top-level directories (Laplace, Levelset, Marmousi, OpenFWI, SmoothVelocity) each contain:
• A Jupyter notebook (.ipynb) orchestrating data loading, model instantiation/training, and result plotting.
• A same-named subfolder capturing artifacts (loss_history.csv, weights-*.hdf5, PDF figures).
• Optional raw data (e.g., Laplace/data/*.txt, Marmousi/vel.rsf).
• requirements.txt lists Python≥3.8 deps (likely TensorFlow/Keras for .hdf5 storage).
• No “service” layers–this is a research pipeline, not a web or microservices app.
Extract Insights from README
• Core components: data ingestion (synthetic/comsol or real velocity fields), model architecture (BranchNet, TrunkNet, plus “root” enrichment net, and add/subtract operations), training engine, evaluation/visualization.
System Components
a. Data Layer
• Raw data stores (txt, rsf) per experiment.
• Preprocessing scripts embedded in notebooks.
b. Model Layer
• En-DeepONet architecture:
• Branch network (parameterizes input function).
• Trunk network (parameterizes spatial coordinates).
• Root network (enrichment).
• Merge operations (add/subtract).
c. Training Pipeline
• Experiment orchestrator (Jupyter notebook + Python scripts).
• Loss computation, optimizer (e.g., Adam/L-BFGS), checkpointing to .hdf5.
d. Postprocessing & Visualization
• Loss history analysis (loss_history.csv → plots).
• Result PDFs (.pdf) and interactive Jupyter outputs.
e. Environment & Dependencies
• Python runtime, Jupyter, TensorFlow/Keras, NumPy, Matplotlib, any comsol-data loader.
Relationships & Data Flows
• Preprocessing → Model inputs.
• Model → Checkpointing (weights-*.hdf5) and LossHistory (.csv).
• Checkpoints & CSV → Postprocessing/Plot module → PDF/Notebook visuals.
Architectural Patterns
• Pipeline pattern: data → model → evaluate
• Modular ML design: separation of “network definition” vs. “training loop” vs. “evaluation scripts.”
• Experiment isolation: each top-level folder is an independent run, sharing common code patterns.
Detail Checklist
• All major folders (Laplace, Levelset, Marmousi, OpenFWI, SmoothVelocity) shown as parallel workflows.
• Data ingestion blocks connected to their respective notebooks.
• Model definition block subdivided into three sub-components (BranchNet, TrunkNet, RootNet) with merge operations highlighted.
• Training block showing optimizer, loss computation → HDF5 & CSV outputs.
• Postprocessing block reading CSV/HDF5 → generating PDF plots.