WSNet: A Deep Learning Library for Engineering Surrogate Modeling

WSNet is an integrated deep learning library specifically designed for high-fidelity surrogate modeling in engineering applications. It provides a unified pipeline for fluid dynamics emulation, structural analysis, and design optimization with comprehensive support for classical surrogate models, neural networks, and modern neural operator algorithms.

🏗 System Architecture

WSNet features a completely reorganized architecture with clear separation of concerns:

wsnet/
├── models/        # Surrogate models (classical, neural, multi-fidelity, ensemble)
├── training/      # Training frameworks and utilities
├── data/          # Data loading and preprocessing
├── sampling/      # Design of Experiments and infill strategies
└── utils/         # Core utilities

1. Models (models/)

A modular repository of surrogate models categorized by their mathematical formulation:

• classical/: Classical response surface algorithms
  • Includes: PRS (Polynomial Response Surface), RBF (Radial Basis Function), KRG (Kriging), SVR (Support Vector Regression)
• neural/: Neural network models
  • Includes: MLP (Multi-Layer Perceptron), DeepONet, GeoFNO (Geometry-aware Fourier Neural Operator), HyperFlowNet, Transolver
• multi_fidelity/: Multi-fidelity models
  • Includes: CCA-MFS, MFS-MLS, MMFS
• ensemble/: Ensemble models
  • Includes: T-AHS, AES-MSI

2. Training (training/)

Training frameworks and utilities:

• base_trainer.py: Base trainer class for custom training workflows
• std_trainer.py: Standard trainer for static regression tasks
• rollout_trainer.py: Trainer for autoregressive sequence prediction
• teacher_forcing_trainer.py: Teacher forcing trainer for sequence prediction
• base_criterion.py: Loss functions and evaluation metrics

3. Data (data/)

Data loading and preprocessing utilities:

• flow_data.py: CFD data loading and preprocessing
• scaler.py: Data scaling utilities (StandardScaler, MinMaxScaler)
• boundary.py: Boundary condition detection and enforcement
• flow_vis.py: Flow field visualization and animation
• flow_plot.py: Training curves, error heatmaps, and metrics plots

4. Sampling (sampling/)

Design of Experiments and infill strategies:

• doe.py: Design of Experiments (LHS, optimized LHS)
• base_infill.py: Base infill sampling strategy
• so_infill.py: Single-objective infill strategy (Expected Improvement)
• mo_infill.py: Multi-objective infill strategy
• mf_infill.py: Multi-fidelity infill strategy

5. Utilities (utils/)

Core utilities and helper functions:

• seeder.py: Reproducibility utilities (seed everything)
• hue_logger.py: Colored logging with ANSI formatting
• sweep.py: Hyperparameter sweep utilities

🚀 Key Features

• CFD-Ready Pipeline: Direct ingestion of ANSYS Fluent data with automatic coordinate and field mapping
• Physics-Informed Training: Support for physics constraints and loss functions
• Multi-Fidelity Support: Comprehensive multi-fidelity modeling capabilities
• Ensemble Methods: Advanced ensemble techniques for improved accuracy
• Neural Operator Algorithms: Modern neural operator implementations for complex physics
• Comprehensive Visualization Tools: Built-in CFD visualization and rendering
• Standardized API: Consistent "initialize, fit, predict" pattern across all models

📚 API Reference

Core Classes

• wsnet.training.base_trainer.Trainer: Base trainer interface
• wsnet.data.flow_data.FlowData: CFD data loader
• wsnet.sampling.doe.lhs_design: Optimized Latin Hypercube Sampling

Common Patterns

The library follows a consistent "initialize, fit, predict" pattern:

# Initialize
model = SomeModel(parameters)

# Fit/Train
trainer = SomeTrainer(model=model)
trainer.fit(train_data, val_data)

# Predict
predictions = model.predict(test_data)

🚀 Examples and Tutorials

Check the examples/ directory for complete workflow examples:

• flow_config.py + flow_train.py: CFD emulation pipeline (HyperFlowNet / GeoFNO / Transolver) with training, inference, and visualization
• aero_config.py + aero_demo.py: Aerodynamic optimization benchmark platform with ensemble, multi-fidelity, sequential sampling, and optimization demos

⚙️ Installation and Dependencies

Requirements

• Python 3.8+
• PyTorch 1.10+
• NumPy
• SciPy
• Matplotlib
• PiVista
• tqdm

Installation

git clone https://github.com/SN-WANG/wsnet.git
cd wsnet
pip install -e .

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

📞 Contact

For questions and support, please contact:

• Shengning Wang (王晟宁) - snwang2023@163.com
• Project Website: https://github.com/SN-WANG/wsnet

📖 Citation

If you use WSNet in your research, please cite:

@software{wsnet2026,
  author = {Shengning Wang},
  title = {WSNet: A Deep Learning Library for Engineering Surrogate Modeling},
  year = {2026},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/SN-WANG/wsnet}}
}