CatDegUS

Python module for Catalysts' Degradation navigated by Uncertainty Sampling. Given a preprocessed time-on-stream catalyst testing data (training dataset), and definitions for target metric (output) and experimental variables (input; range and step size), python codes based on CatDegUS can suggest the most informative experimental condition with the largest uncertainty. 2D/3D distribution of the GP-based uncertainty can also be visualized.

Getting started

1. Make a virtual environment (e.g., when using conda):

conda create -n catdegus python=3.13
conda activate catdegus

2. Installation

• Directly install using pip (the simplest way)

  pip install git+https://github.com/dongjae-shin/CatDegUS.git

• cf. Alternative way: clone repository & install using pip

  git clone https://github.com/dongjae-shin/CatDegUS.git
  cd CatDegUS
  pip install .

• cf. You can also install all the requirements by:

  pip install -r requirements.txt

3. Run example codes

• In the tests/, run as follows:

  python ./example.py

• Alternatively, in a Jupyter notebook (example.ipynb):

  import catdegus.active_learning.gaussian_process as gpc
  
  # Define the home directory and path to data
  # Target metric: initial CO2 conversion
  path = "./20250228_sheet_for_ML_unique.xlsx"
  
  # Train the Gaussian Process model
  GP = gpc.GaussianProcess()
  GP.preprocess_data_at_once(path=path,
                            target='CO2 Conversion (%)_initial value',
                            x_range_min=[300, 0.1, 0.005, 0], 
                            x_range_max=[550, 1.0, 0.02, 1])
  GP.train_gp()

Example codes:

• Example python codes to use CatDegUS are in tests/ directory.
• example.ipynb (or example.py): sequential uncertainty sampling for catalyst testing and the visualization of uncertainty.
• example_HT_reactor: batch uncertainty sampling for high-throughput (HT) reactor with specific 4×4 reactor architecture.

Requirements

• Required modules: pandas, torch, botorch, matplotlib, openpyxl
• All specified in setup.py

Supported acquisition functions

• Posterior Standard Deviation: used for uncertainty sampling (US)
• Posterior Mean
• Upper Confidence Bound (UCB)
• Expected Improvement (EI): to be added.

Input to the code

• Path to a data file (*.xlsx): example
• Target metric as output of GP surrogate model, e.g., ‘CO2 Conversion(%)_initial value’; you can choose one of target column names
• Lower/Upper boundaries for input features (reaction temperature, Rh weight loading, Rh total mass, and synthesis method), by which the search space is bounded
• Step sizes for input features, which are allowed by experimental resolution
• Number of conditions to suggest; top $n$ informative conditions will be suggested.
• Temperatures to plot 2D acquisitions functions for
• Synthesis method 1) to suggest the most informative experimental condition for and 2) to plot acquisition functions for

Output from the code

Uncertainty Sampling

• Maximizer condition for posterior standard deviation: US-guided experimental condition
  
• Maximizer condition for other supported acquisition functions
• Joint maximizer condition for posterior standard deviation for q-batch sampling
• Selection of temperatures with uncertainties averaged over the other features for specific HT reactor

Visualization

• 2D visualization of a selected acquisition function for a selected synthesis method and temperature
  
• 3D visualization of a selected acquisition function for a selected synthesis method
  

Potential Integration with Other Models

• JSON output file corresponding to 2D/3D plot above is generated upon visualization, which could be used for applications including LLM-agent.
  

Related Publication

• Integration into a LLM-based multi-agent system (link)

To do

• Extension of available acquisition functions
• SHAP analysis
• Selection of the best regression model by LOOCV score
• Web application for non-experts