dklbo

Overview

dklbo is a software to perform Bayesian optimization (BO) using Deep Kernel Learning (DKL) for materials search.
This software is not directly applicable to individual cases because the neural network part of DKL depends on the specific task.

Therefore, this package serves as an example to demonstrate how to use DKL.

Instructions

This package assumes the use of graph-based neural networks and modifies 's ExactGP class in gpytorch to create a custom class ExactGP_graph, which allows Data class in pytorch_geometric to be used as input.

Additionally, to facilitate handling high-entropy alloys with multiple sites, a custom data structure called SiteGraph, which inherits from Data class in pytorch_geometric, is used.
This enables handling multiple sites individually within a single Data object.

The main modules are in the dkl.py. The DKL class defines the architecture of DKL.
By receiving Data in the forward function, it performs feature transformation through the neural network component, then inputs the transformed features into Gaussian process regression, and outputs the predictive mean and variance.

In the DKL class, transformer (the neural network part of DKL), as well as the Gaussian process components, mean_module and covar_module, can be passed as arguments when instantiating the class.
Alternatively, they can be defined within __init__ to create a new class, as seen in DKLComGraph and EDKLCGCNN.

Sample Code

1. /example/search_bandgap.py

• Unzip /example/datasets/calculation/cifs.zip
• Searches for the material with the largest band gap among 922 oxides.
• Uses CGCNN for the neural network.
  • 📄 Reference: CGCNN

python search_bandgap.py

2. /example/search_expbandgap.py

• Searches for the material with the largest band gap among 610 organic-inorganic hybrid perovskites (ABX₃).
• Uses a graph-based neural network, where convolution is applied to each A-site, B-site, and X-site separately.
• The SiteGraph structure is utilized to manage site-specific information in ABX₃.

python search_expbandgap.py

Installation

# Clone the repository
git clone https://github.com/skiyohara/dklbo.git

# Required packages
pytorch
pytorch_geometric
pytorch_scatter
gpytorch
pymatgen