StressNet: Deep Learning for Structural Stress Prediction

An AI-driven structural analysis project that leverages Machine Learning and Deep Learning to predict the maximum Von Mises stress of mechanical components from Finite Element Analysis (FEA) simulation data. The project compares classical regression techniques with neural networks to provide accurate stress predictions while reducing dependence on computationally expensive simulations.

---

Overview

Finite Element Analysis (FEA) is widely used to evaluate structural performance under different loading conditions. However, running large-scale FEA simulations can be computationally expensive.

StressNet aims to learn the relationship between:

• Structural geometry
• Thickness parameters
• External loading conditions

and predict the maximum Von Mises stress directly using data-driven models.

---

Features

• Automated preprocessing of FEA simulation datasets

• Feature engineering using:

  • Thickness parameters
  • Node coordinates
  • External forces

• Linear Regression baseline

• Polynomial Regression with Ridge Regularization

• Deep Neural Network using TensorFlow/Keras

• Model comparison using:

  • RMSE
  • MAE
  • R² Score

• Training and validation loss visualization

• Actual vs Predicted parity plots

• Saved trained neural network model for inference

---

Dataset

The model is trained using thousands of FEA simulation samples.

Each sample contains:

• Structural thickness values
• Node coordinates
• Applied force vectors
• Maximum Von Mises stress

---

Machine Learning Pipeline

1. Load FEA datasets
2. Extract structural and loading features
3. Preprocess and normalize data
4. Train multiple regression models
5. Train Deep Neural Network
6. Compare model performance
7. Save the best performing model

---

Models Evaluated

• Linear Regression
• Polynomial Regression (Ridge)
• Deep Neural Network (TensorFlow/Keras)

---

Technologies Used

• Python
• TensorFlow / Keras
• Scikit-learn
• NumPy
• Pandas
• Matplotlib
• Finite Element Analysis (FEA)

---

Evaluation Metrics

The models are evaluated using:

• Root Mean Square Error (RMSE)
• Mean Absolute Error (MAE)
• R² Score

to compare prediction accuracy and generalization performance.

---

Results

The Deep Neural Network achieved significantly better predictive performance than traditional regression models, demonstrating its ability to capture complex nonlinear relationships between geometry, loading conditions, and structural stress.

---

Future Improvements

• Physics-Informed Neural Networks (PINNs)
• Graph Neural Networks for mesh-based learning
• Transformer-based surrogate models
• Uncertainty estimation
• Model deployment using FastAPI
• Interactive web dashboard for inference

---

Author

Vansh Garg

IIT Ropar