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 contains both classical response surface algorithms, standard deep learning algorithms and modern neural operator algorithms, providing a unified pipeline for fluid dynamics emulation, structural analysis, and design optimization.

---

🏗 System Architecture

The library is organized into three primary pillars, ensuring a clear separation between algorithms, utilities, and applications.

1. Algorithms (nets/)

A modular repository of algorithms categorized by their mathematical formulation:

• surfaces/: Classical response surface algorithms.
  • Includes: PRS (Polynomial Response Surface), RBF (Radial Basis Function), KRG (Kriging), and SVR (Support Vector Regression).
• baselines/: Standard deep learning algorithms.
  • Includes: MLP (Multi-Layer Perceptron).
• operators/: Modern neural operator algorithms.
  • Includes: DeepONet (Deep Operator Network), GeoFNO (Geometry-aware Fourier Neural Operator).

2. Utilities (utils/)

Production-grade utilities tailored for physical datasets:

• DoE: Implementation of LHS (Latin Hypercube Sampling) with Maximin distance optimization for maximum information gain in parameter Design of Experimnets.
• Engine: Fully encapsulated training pipeline serving as a "One-Stop" solution for DL workflows. It features:
  • TensorScaler: Channel-wise normalization tailored for multi-physics fields.
  • AutoregressiveTrainer: Trainer specialized for sequence rollout with pushforward logic and noise injection to ensure long-term spatio-temporal stability.
• CFDParser: Automated data ETL (Extract, Transform and Load) pipeline for ANSYS Fluent exports. It handles raw .txt exports, supports spatial/temporal subsampling, and implements high-speed .pt caching.
• CFDRender: Automated visualization pipeline for CFD results, generating side-by-side animations of Ground Truth, Prediction, and Absolute Error.

3. Applications (apps/)

End-to-end research workflows demonstrating the application of wsnet to complex physical domains. These applications provide best practices for integrating datasets, models, and trainers to solve production-level problems.

• HyperFlow-Net (Fluid Dynamics Emulation):
  • Application: Real-time emulation of hydrogen energy pipelines driven by extreme high-pressure differentials.
  • Methodology: Utilizes GeoFNO for mesh-independent mapping to capture non-linear shock wave propagation and rapid pressure transients in complex piping topologies.
• FId-Net (Structural Analysis & Inverse Problems):
  • Application: Force Identification (FId) for plate structures.
  • Methodology: A deep learning approach for solving Inverse Problems. It processes multi-sensor vibration/strain data to reconstruct the magnitude and spatial coordinates of external impact loads with high precision.
• AeroOpt-Solver (Aerospace Design Optimization):
  • Application: High-fidelity aerodynamic design optimization for aerospace components.
  • Methodology: Leverages wsnet's surrogate models (KRG/RBF) and DoE utilities to accelerate the optimization loop, significantly reducing the computational cost compared to traditional CFD-based adjoint methods.

---

🚀 Key Features

• CFD-Ready Pipeline: Direct ingestion of Fluent data with automatic coordinate and field mapping (Vx, Vy, Vz, P, T).
• Physics-Inspired Training: The engine supports Curriculum Learning for rollout steps, allowing models to learn short-term dynamics before tackling long-term trajectories.
• Optimization-Driven DoE: Generate space-filling designs using optimized LHS to maximize information gain in the parameter space.
• Quality Inspection: Integrated rendering tools to monitor model performance across multiple physical fields simultaneously.

---

🛠 Quick Start

Data Preparation

Place your ANSYS Fluent .txt exports in the dataset/ folder following the case_0001/ naming convention.

Training an Autoregressive Model

import torch
from wsnet.utils.CFDParser import CFDataset
from wsnet.utils.Engine import AutoregressiveTrainer
from wsnet.nets.operators.GeoFNO import GeoFNO # Example

# 1. Load Data
train_data, val_data, _ = CFDataset.build_datasets(data_dir='./dataset', spatial_dim=2)

# 2. Initialize Model & Trainer
model = GeoFNO(...)
trainer = AutoregressiveTrainer(
    model=model,
    output_dir='./runs'
    pushforward_steps=5, 
    noise_std=0.005,
)

# 3. Fit
trainer.fit(train_loader, val_loader)