eVTOL-VehicleModel 🚀

Overview

The aerodynamic performance of a multirotor eVTOL aircraft is strongly influenced by the complex interactions between rotors and wings. Due to the non-linear and unsteady nature of the flow, traditional aerodynamic modeling approaches rely on high-fidelity CFD simulations, which are computationally expensive.

This repository presents a data-driven surrogate aerodynamic model based on Graph Neural Networks (GNNs) to predict the aerodynamic coefficients of aircraft components. Two distinct graph-based architectures—Hierarchical and Recurrent—are implemented to model the rotor-rotor and rotor-wing interactions. The proposed framework integrates deep learning techniques to streamline the aerodynamic modeling process, making it more computationally efficient for the conceptual design phase.

---

Features

Modular Aerodynamic Modeling – Airfoil, wing, and rotor models are trained separately.
Graph Neural Network (GNN) – Captures rotor-to-rotor and rotor-wing interactions.
Hierarchical vs. Recurrent Architectures – Two different graph-based learning strategies.
CFD-Based Training Data – Uses CFD simulation data from FLOWUnsteady.
Optimized Deep Learning Models – CNNs, LSTMs, and GNNs tailored for aerodynamic predictions.

---

Models & Data

The surrogate aerodynamic model is developed in a modular way, consisting of the following components:

1 Airfoil Model

• Architecture: Convolutional Neural Network (CNN)
• Objective: Predict lift (Cl) and drag (Cd) coefficients from airfoil coordinates.
• Training Data: XFOIL-based aerodynamic coefficients.

2 Wing & Rotor Models

• Architecture: Long Short-Term Memory (LSTM)
• Objective: Predict aerodynamic performance coefficients (Cl, Cd, Ct, Cq) for wings and rotors.
• Training Data: FLOWUnsteady simulation results incorporating variations in rotor speeds, angles of attack, and freestream velocity.

3 Composite Model (eVTOL Aircraft)

• Architecture: Graph Neural Network (GNN)
• Objective: Predict the aerodynamic coefficients of the full aircraft by integrating rotor and wing models.
• Graph Edges: Spatial relationships between aircraft components to model rotor-wing and rotor-rotor interactions.
• Graph Architectures:
  • Hierarchical GNN – Modular and interpretable training.
  • Recurrent GNN – Captures sequential dependencies but is computationally heavier.
• Performance Comparison: The Hierarchical GNN achieves an average relative L2 error of 5%, outperforming the recurrent model in accuracy and computational efficiency.

---

Usage

1 Utility Functions & Model Architectures

• Located in /src/
• Includes Python scripts for data processing, and model architecture.

2 Jupyter Notebooks for Training & Evaluation

• Located in /notebooks/
• Notebooks with *_analysis.ipynb:
  • Load trained models (airfoil, rotor, wing, aircraft).
  • Plot results and performance metrics.

3 Pre-trained Models & Scalers

• Located in /trained_models/
• Contains saved models for airfoil, wing, rotor, and composite GNN models.

4 Model Performance & Results

• Located in /results/
• Stores plots and comparisons between Hierarchical vs. Recurrent GNN architectures.

5 Training & Test Data

• Located in /data/
• Includes the dataset used for training and evaluating the models.

---

Installation

1 Prerequisites

This system requires Windows Subsystem for Linux (WSL) to run properly.
Ensure that WSL 2 is installed and set up before proceeding.

2 Clone the Repository

git clone https://github.com/ss-hegde/eVTOL-VehicleModel.git

cd eVTOL-VehicleModel

pip install -r requirements.txt

Contact

For questions or feedback, reach out via: email: [email protected]