BALPI

This repository contains an implementation of BALPI, a Bayesian active learning framework for phase identification and phase-diagram exploration.

BALPI studies phase identification under two related formulations:

• Classification, where the target is represented as a discrete phase label.
• Level-set estimation / regression, where a continuous phase score or phase fraction is queried and a phase boundary is recovered by thresholding.

The implementation uses Gaussian-process surrogate models and acquisition utilities to adaptively select new compositions to query.

Paper

This code accompanies:

Bayesian Active Learning to Accelerate High Throughput Phase Diagram Exploration

If you use this repository, please cite the paper:

@article{fan2026balpi,
  title={Bayesian Active Learning to Accelerate High Throughput Phase Diagram Exploration},
  author={Fan, Mingzhou and Wang, Yucheng and Vazquez, Guillermo and Zhou, Ruida and Karaman, Ibrahim and Arroyave, Raymundo and Qian, Xiaoning},
  year={2026}
}

Repository Contents

BALPI/
+-- code/
|   +-- dataset.py
|   +-- experiment.py
|   +-- main.py
|   +-- optimization.py
|   +-- surrogate.py
|   +-- util.py
|   +-- utilityfunction.py
+-- data/
|   +-- toy_data.csv
|   +-- toy_data_regression.csv
+-- Flow_Chart.png
+-- LICENSE
+-- README.md
+-- pyproject.toml
+-- requirements.txt

The main code components are:

• dataset.py: CSV-backed data interface and toy data utilities.
• surrogate.py: Gaussian-process classifier/regressor and baseline models.
• utilityfunction.py: acquisition functions, including MES, BALD, SMOCU, and straddle/UCB-style utilities.
• optimization.py: Monte Carlo acquisition maximization.
• experiment.py: active-learning experiment classes.
• main.py: example regression / level-set active-learning run.

Data

The bundled data/ directory contains the BCC-B2 NiTiHfCu pseudo-ternary data used for the target maps in Figure 4(a-b) of the paper.

• data/toy_data_regression.csv contains a continuous BCC-B2 phase score / phase-fraction map. This corresponds to Figure 4(a).
• data/toy_data.csv contains the corresponding binary BCC-B2 classification map. It is obtained by thresholding toy_data_regression.csv at 0.8, and corresponds to Figure 4(b).

Both CSV files use the same format:

x1,x2,x3,target

where x1, x2, and x3 are ternary composition coordinates satisfying x1 + x2 + x3 = 1, and target is either a continuous score or a binary label.

The Figure 3 data are not redistributed in this repository. The source data for the SiO2-Al2O3-MgO glass-ceramic phase-identification example can be found in:

M. Lesniak, J. Partyka, K. Pasiut, M. Sitarz, Microstructure study of opaque glazes from SiO2-Al2O3-MgO-K2O-Na2O system by variable molar ratio of SiO2/Al2O3 by FTIR and Raman spectroscopy, Journal of Molecular Structure 1126 (2016) 240-250.

Installation

The project was developed against older scientific Python packages. A Python 3.8 environment is recommended.

Using venv:

python3.8 -m venv .venv
source .venv/bin/activate
pip install --upgrade pip setuptools wheel
pip install -r requirements.txt

Using conda:

conda create -n balpi python=3.8
conda activate balpi
pip install --upgrade pip setuptools wheel
pip install -r requirements.txt

Running the Example

From the repository root:

python code/main.py

The example in main.py:

• loads data/toy_data_regression.csv,
• initializes a Gaussian-process regression active-learning experiment,
• uses a straddle/UCB-style acquisition utility,
• runs repeated active-learning iterations, and
• writes outputs to a directory named like results_BCC_20/.

By default, main.py uses:

• init = 20 initial samples,
• iters = 80 active-learning iterations,
• 10 repeated runs,
• utilityfunction.U_UCBS(x, model, .8, .5), and
• the threshold 0.8 for identifying the BCC-B2 region.

This run can be computationally expensive because it repeatedly retrains Gaussian-process models. For a quick smoke test, reduce init, iters, the outer repeat loop, or the Monte Carlo search size in main.py.

License

See LICENSE.