RHEA-Inverse-Design-Using-CVAE

This repository implements a Conditional Variational Autoencoder (CVAE) for the inverse design of Refractory High Entropy Alloys (RHEAs), focusing on predicting and generating candidate alloys with optimized yield strength under varying testing temperatures.

---

Project Overview

• Domain: Materials Informatics, Inverse Design, Deep Generative Models
• Objective: Generate new alloy compositions with desired mechanical properties (e.g., yield strength at high temperatures).
• Approach:
  1. Encode alloy compositions, processing conditions, and properties into latent space.
  2. Train a CVAE conditioned on temperature.
  3. Generate candidate alloys matching target yield strength.
  4. Validate with explainability analysis (correlation + permutation feature importance).

---

Repository Structure

RHEA-Inverse-Design-Using-VAE/
│
├── data/
│   ├── scripts/rhea_data_encoding.py     # Data cleaning & encoding
│   ├── data.csv                          # Raw dataset
│   ├── encoded_data.csv                  # Cleaned + encoded dataset (model input)
│   └── processed/                        # Processed splits + scalers
│
├── models/
│   └── cvae_best.pt                      # Trained model checkpoint
│
├── outputs/                              # Generated alloys, plots, explainability results
│   ├── explainability/
│       ├── correlation_train_vs_generated.png
│       └── pfi_importance.png
│   └── tsne/
│       ├── latent_tsne_all.png
│       ├── latent_tsne_yield_strength.png
│       └── latent_tsne_temperature.png
│
├── src/
│   ├── data_prep.py              # Train/val split, scaling
│   ├── cvae.py                   # CVAE model definition
│   ├── train_cvae.py             # Training script with early stopping
│   ├── generate.py               # Alloy generation (sampling + refinement)
│   ├── evaluate_cvae.py          # Interactive query interface
│   ├── latent_vis.py             # Latent space visualization (t-SNE)
│   └── explainability.py         # Correlation + PFI explainability
│
└── README.md

---

Installation

git clone <https://github.com/shruti-sivakumar/RHEA-Inverse-Design-Using-VAE>
cd RHEA-Inverse-Design-Using-VAE
pip install -r requirements.txt

Requirements:

• Python 3.9+
• PyTorch
• NumPy, Pandas, Scikit-learn
• Seaborn, Matplotlib
• Joblib

---

Usage

1. Data Cleaning & Encoding

python data/scripts/rhea_data_encoding.py

Generates:

• encoded_data.csv (model-ready, numeric)
• encoded_data_human.csv (readable, for analysis)

2. Preprocessing

python src/data_prep.py

Outputs scalers + splits in data/processed/.

3. Train CVAE

python src/train_cvae.py

• Early stopping enabled
• Saves best model to models/cvae_best.pt

4. Generate Candidate Alloys

python src/evaluate_cvae.py

Interactive menu with 3 query modes:

1. Highest yield strength at given temperature
2. Closest to a target yield strength
3. Multi-constraint (two temperature constraints)

5. Latent Space Visualization

python src/latent_vis.py

• t-SNE plots of latent distribution
• Colored by yield strength & temperature

6. Explainability

python src/explainability.py

Produces:

• Correlation heatmap: Train vs Generated alloys
• Permutation Feature Importance (PFI) plot

---

Results & Visualizations

Latent Space (t-SNE)

Explainability

Correlation Heatmap (Train vs Generated):

Permutation Feature Importance (PFI):

---

Key Insights

• CVAE can successfully generate novel alloy compositions with desired yield strength.
• Strong conditioning ensures properties are tuned for different testing temperatures.
• Explainability confirms both data fidelity (correlation analysis) and model interpretability (PFI).

---

Next Steps

• Add case-study Integrated Gradients (IG) for per-alloy local explanations.
• Benchmark against other generative models (GANs, diffusion models).
• Validate generated alloys with external simulation/experimental datasets.

---

License

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