GNN-Frequency-Dependent-Optical-Properties

This project leverages Graph Neural Networks (GNNs) to predict the frequency-dependent dielectric function from crystal structures. The data and model provided in this repository support the findings from the paper:
"Prediction of Frequency-Dependent Optical Spectrum for Solid Materials: A Multioutput and Multifidelity Machine Learning Approach"
Read the paper here.

Project Overview

The repository provides code and data for training and evaluating a GNN-based model that predicts the optical spectra of materials across various frequencies. The approach integrates multi-fidelity data and compares multi-output scaling/loss architectures to enhance the learning process of high-fidelity optical spectra.

Key Features

• Multi-Fidelity Learning: Employs and compares fidelity embeddings and transfer learning to integrate data from both high- and low-fidelity sources.
• Custom GNN Architecture: Implements a custom architecture with MEGNet layers for effective feature extraction from crystal structures.
• Custom Loss Functions: Applies and compares tailored loss functions (MAE, KL divergence, Wasserstein) and scaling schemes (UnNorm, MaxNorm, AvgNorm) to minimize prediction errors for optical spectra.

How to Use

Environment Setup

Ensure you have the following dependencies installed:

• Python 3.x
• TensorFlow 2.x
• MEGNet
• Pymatgen/ASE (for structure handling)
• Optuna (for hyperparameter optimization)
• Seaborn, Matplotlib (for plotting)

Model Training

Modify the input parameters/hyperparameters directly in the code as needed, and execute it to start training:

• loss_weights: Adjust in case of normalized spectrum learning.
• embedding_dimensions for atom, bond, and state features.
• GNN layer sizes, Dropout rates, and Learning rate, etc.

Model Evaluation

The script automatically evaluates the model using a test set, plotting both training and validation losses over epochs. It also calculates custom metrics for component-wise errors across the predicted spectra and target norms.

General Layout

GNN Predictions for the Imaginary Part of the Dielectric Function and the Absorption Coefficient

Predictions of the imaginary part of the dielectric function and the absorption coefficient at meta-GGA MBJ functional accuracy.

GNN Predictions for the Short-Circuit Current, Reverse Saturation Current, and Spectroscopic Limit of Maximum Efficiency