FABO: Feature-Adaptive Bayesian Optimization Framework

Overview

FABO (Feature-Adaptive Bayesian Optimization) is an advanced framework designed to enhance material discovery by dynamically selecting the most relevant features at each iteration of the optimization process. This novel integration of feature selection techniques within the Bayesian Optimization (BO) loop ensures reduced data dimensionality and increased efficiency, especially in high-dimensional search spaces.

This repository provides the code and resources to run FABO on tasks such as CO₂ uptake and electronic property optimization for MOF discovery. FABO adapts to data distributions, making it more flexible than traditional BO methods reliant on static feature sets.

---

Key Features

• Dynamic Feature Selection: Updates feature sets iteratively to improve optimization efficiency.
• Flexible Acquisition Strategies: Supports Expected Improvement (EI) and Upper Confidence Bound (UCB), enabling better exploration and exploitation.
• Robust Surrogate Model: Employs a Gaussian Process to predict outcomes and quantify uncertainty.
• Task-Agnostic: Applicable to diverse optimization problems, including adsorption and electronic property optimization.

---

Installation

Clone the repository and install the required dependencies:

git clone https://github.com/AI4ChemS/FABO.git
cd FABO
pip install -r requirements.txt

---

Customizing FABO for Your Application

FABO is highly customizable. In the FABO_main notebook, you can adjust parameters to align with your specific optimization needs:

• nb_iterations: Number of Bayesian optimization iterations.
• nb_initialization: Number of random initial samples for optimization.
• min_features and max_features: Range of features selected per iteration (default: 5 to 20).
• FABO: Boolean flag to enable (True) or disable (False) Feature-Adaptive Bayesian Optimization.
• which_acquisition: Acquisition function choice: "EI", "max y_hat", or "max sigma".
• FS_method: Feature selection method ('spearman' or 'mRMR').
• n_seed and seeds: Control randomness in optimization (default: 19 seeds).
• path: Path to the dataset (CSV format).
• label: Target property for optimization (e.g., "pure_uptake_CO2_298.00_1600000").

Example Usage

To optimize a new task, modify these parameters:

1. path: Set to your dataset file path (e.g., "data/my_dataset.csv").
2. label: Define the property to optimize (e.g., "target_property").
3. which_acquisition: Choose based on your preference for exploration ("max sigma") or exploitation ("max y_hat").
4. min_features and max_features: Adjust to your task's feature range.

By tailoring these settings, FABO can adapt to efficiently optimize material discovery tasks.

---

Feel free to reach out if you have questions or need assistance!