A partially quantum‑informed crystal graph network for direct adsorption energy prediction in catalysis

Area of Science:

• Materials Science
• Computational Chemistry
• Catalysis

Background:

• Rational catalyst design needs accurate yet efficient predictive tools for adsorbate energetics.
• Current machine learning models often rely on computationally intensive methods like total energy decomposition and force supervision.

Purpose of the Study:

• To introduce Q-CatNet, a novel machine learning model for direct prediction of adsorption energy.
• To bypass computationally expensive methods by directly correlating initial structure to adsorption energy.

Main Methods:

• Q-CatNet extends the Crystal Graph Convolutional Neural Network (CGCNN) framework.
• Incorporates quantum-informed edge descriptors (electrostatic Hamiltonian) and global features (DOS, XRD).
• Directly predicts structure-to-adsorption energy for various adsorbates on catalyst surfaces.

Main Results:

• Q-CatNet demonstrates robust performance on a curated dataset.
• Outperforms image-based Fourier-Transformed Crystal Property (FTCP) by 46%.
• Exceeds several invariant graph-based architectures by 8% to 38%.

Conclusions:

• Q-CatNet provides a physically grounded framework for adsorption energy prediction.
• Enables faster and more practical predictions, accelerating catalyst discovery.
• Bypasses the need for total energy decomposition, improving computational efficiency.