Deep Learning Accelerated Multi-Criteria Screening of Chromium-Based A2+B2 Superalloys Across 11 Elements

This repository contains datasets and scripts to reproduce selected datasets and results of the article 'Accelerated discovery of Cr-based A2+B2 superalloys across 11 elements with a deep-learning CALPHAD surrogate'.

Prerequisites

A requirements.txt file is included for installing the required Python packages (excluding tc-python, which must be installed after Thermo-Calc).

System Requirements

• Python 3.9.17
• Thermo-Calc version 2025b

Python Packages

Install the required packages using:

pip install -r requirements.txt

Thermo-Calc Setup

1. Install Thermo-Calc (tested with version 2025b).
2. Install tc-python (tc-python==2025.2.30) using the Python wheel provided with Thermo-Calc 2025b.

Note: As Thermo-Calc licenses do not permit the open release of large calculated datasets, only a 100-composition 'dummy' version of the full set of feasible compositions (roughly 15,000 compositions) is provided. Therefore, to reproduce some of the figures in the manuscript, a license to the commercial TCHEA7 database is needed.

Repository Structure

• data_generation/: Scripts for generating and processing CALPHAD datasets
• surrogate_modeling/: Code for training surrogate models
• screening/: Screening pipeline for candidate compositions
• analyze_feasible_candidates/: Analysis and visualization of screening results
• analyze_experimental_compositions/: Analysis of experimental data
• helper_functions/: Utility functions for calculations (yield strength, VEC, misfit volumes)
• shared/: Shared utilities and common functions

Workflow

Follow these steps to reproduce the datasets and results:

1. Generate Composition Pool for Generating Training Data

Run the data_generation/training_pool_generator.py to create composition pool compositions_b.h5.

2. Generate Dataset B

Run the data_generation/calphad_data_generator.py to reproduce Dataset B dataset_b.h5 using compositions_b.h5 as input.

Note: Due to CALPHAD convergence issues we iteratively checked which compositions were missing from the dataset, and repeated calculations for these compositions until dataset_b.h5 included 99.98% (3,570,288) of the CALPHAD data points.

3. Postprocess and Split Dataset B

Run the data_generation/postprocess_and_split_dataset.py to process and split dataset_b.h5 into development and test sets. This creates dataset_b_dev.h5 and dataset_b_test.h5.

4. Train Surrogate Model

Run surrogate_modeling/main.py to create the surrogate model with optimized hyperparameters. Move the generated model/ folder to the screening/ folder.

5. Generate Composition Pool

Create the composition candidate pool for screening with data_generation/screening_pool_generator.py. Move the generated composition_pool_screening.h5 to the screening/ folder.

6. Perform Screening

Execute the screening step with screening/main.py to identify feasible candidates. This produces feasible_candidates.xlsx.

7. Analyze Results

-analyze_feasible_candidates/: analyze the feasible compositions (alloy families, VEC-strength scatter plots). -analyze_experimental_compositions: contains the scripts to visualize DFT-derived properties (strength, D-parameters).

Usage Tips

• Ensure Thermo-Calc is properly configured before running data generation scripts.
• When working with CALPHAD data, be aware of potential convergence issues and missing data points.
• Check file paths and dependencies when moving generated files between folders

License

This project is licensed under the GNU Affero General Public License v3.0 or later (AGPL-3.0-or-later). See the LICENSE file for details.

Citation

• TBA