Deepcoupling

<h1 align="center">Accelerated Photocatalytic C-C Coupling via Interpretable Deep Learning: Single-Crystal Perovskite Catalyst Design Using First-Principles Calculations</h1> <p align="center">Yuze Hao, Lan Duo</p> <p align="center">College of Chemistry and Chemical Engineering, Inner Mongolia University, Hohhot 010021, China.</p> <p align="center">Photocatalytic C–C coupling reactions have garnered significant attention for their potential to drive sustainable chemical transformations. The design of efficient photocatalysts is critical in optimizing these reactions. In this study, we use a computational materials science approach, leveraging first-principles calculations to evaluate the bandgap values of 158 single-crystal perovskite materials. We employ a deep learning model, incorporating a multi-head-attention mechanism within a ResNet architecture, to predict the bandgap based on features such as τ , Group-A, Group-B, Pettifor number, χM-B, χP-B, Ea-A, cB, KB, and Ra-B. This model’s performance is compared to traditional machine learning techniques, including K-means, MLP, Random Forest, PCA, and Multivariable Linear Regression. The results demonstrate that the self-attention ResNet model achieves a training R2 of 0.819 and a test R2 of 0.803, indicating strong predictive accuracy. The model’s interpretability is enhanced by visualizing the permutation importance of each feature, shedding light on the contributions of various factors to the prediction. These findings highlight the potential of machine learning, particularly deep learning, in accelerating the design of photocatalysts for C–C coupling reactions.</p> <p align="center"></p>

Table of Contents

• Project Overview
• Key Features
• Data Description
• Model Architecture
• Results and Performance
• Installation
• Usage
• Citation

---

Project Overview

Photocatalytic C-C coupling reactions are essential for sustainable chemical synthesis. This research utilizes a computational approach, leveraging first-principles calculations to evaluate 158 single-crystal perovskite materials. By integrating a multi-head-attention mechanism with a ResNet architecture, we provide a robust tool for predicting bandgap values, which are critical for catalyst efficiency.

Key Features

• Deep Learning Integration: Combines ResNet's residual learning with Multi-head Self-Attention for feature weighing.
• Interpretable AI: Focuses on 10 key physical and chemical descriptors selected from a pool of 38 features.
• High Performance: Outperforms traditional methods like Random Forest, MLP, and PCA in bandgap prediction tasks.

Data Description

The dataset consists of 158 single-crystal perovskite materials. Feature selection was performed using Pearson correlation analysis to identify the most impactful descriptors.

Selected Descriptors

1. Transition Metal Parameter ($\eta$)
2. Group-A / Group-B (Element groupings)
3. Pettifor Number
4. $\chi M-B$ / $\chi P-B$ (Electronegativity)
5. $E_a-A$ (Activation Energy)
6. $C_B$ (Bonding Parameter)
7. $K_B$ (Bulk Modulus)
8. $R_a-B$ (Atomic Radii)

Figure 1: Heatmap of Pearson correlation analysis for feature selection.

Model Architecture

The model utilizes a 10-layer ResNet backbone enhanced with a Multi-head Attention layer to capture complex inter-dependencies between the input descriptors.

Figure 2: Schematic of the Multi-head-Attention enhanced ResNet architecture.

Results and Performance

Our model demonstrates superior stability and predictive power compared to traditional regression and machine learning techniques.

| Model | Training $R^2$ | Test $R^2$ | | :--- | :---: | :---: | | Proposed ResNet-Attention | 0.819 | 0.803 | | Random Forest | 0.946 | 0.581 | | MLP (Multilayer Perceptron) | 0.792 | -66.437 | | K-means | 0.583 | 0.031 | | PCA | 0.439 | -0.10 |

Figure 3: Predicted vs. Actual bandgap values for the test set.

Installation

1. Clone the repository:

   git clone [https://github.com/username/project-name.git](https://github.com/username/project-name.git)
   cd project-name
   

2. Install dependencies:

   pip install -r requirements.txt
   

Usage

Training

To train the model from scratch using the provided dataset:

python train.py --data data/perovskite_data.csv --epochs 100

Prediction

To use a pre-trained model for bandgap prediction:

python predict.py --input sample_descriptors.json

Citation

If you use this code or the findings in your research, please cite:

@article{hao2025accelerated,
  title={Accelerated Photocatalytic C-C Coupling via Interpretable Deep Learning: Single-Crystal Perovskite Catalyst Design Using First-Principles Calculations},
  author={Hao, Yuze and Duo, Lan},
  journal={ICLR 2025},
  year={2025}
}