ML-YS-MPEAs

A machine learning framework for training artificial neural networks to predict yield strength of multi-principal element alloys (MPEAs) using cross-validation and automated feature engineering.

---

Publication details:

Machine learning driven insights into yield strength of multi-principal element alloys

• Authors : Dishant Beniwal ¹, Ashish Kaushik ¹, Abhishek Tiwari ¹, Pratik K. Ray ¹
• Journal : Journal of Materials Research
• DOI : 10.1557/s43578-025-01753-x

---

The pre-trained model for Yield strength will soon be integrated in MAPAL alongside the already existing phase selection and hardness prediction models.

---

Features

• Automated Neural Network Training: Build and train ANN models with customizable architectures • Cross-Validation: K-fold cross-validation with multiple runs for robust model evaluation • Feature Engineering: Automatic calculation of alloy features including:

• Elemental properties (mean, variance, etc.)
• Miedema enthalpy calculations (H_chem, H_el, H_mix, H_IM)
• Phase fraction predictions (f_FCC, f_BCC, f_IM) • Model Persistence: Automatic saving of trained models, predictions, and performance metrics • Convergence Monitoring: Ensures models converge properly with error threshold checking

Dependencies

Package	Purpose
pandas	Data manipulation and analysis
numpy	Numerical computing
tensorflow	Neural network framework
scikit-learn	Machine learning utilities
mapal	Alloy featurization and pre-trained model transfer learning
argparse	Command line argument parsing
openpyxl	Excel file reading
xlsxwriter	Excel file writing with multi-sheet support

Installation

• Clone this repository • Install required dependencies:

pip install pandas numpy tensorflow scikit-learn openpyxl xlsxwriter mapal

Usage

Command Line Arguments

Argument	Type	Description	Required
-input_filepath	String	Path to the Excel input file containing model configurations	Yes
-modIdStart	Integer	Starting model ID from the input file	Yes
-modIdEnd	Integer	Ending model ID from the input file	Yes
-saveDir	String	Directory to save trained models	No (default: "trained-models-raw")

Basic Usage Examples

• Single model training:

python run-autoANN.py -input_filepath models_config.xlsx -modIdStart 1 -modIdEnd 1

• Multiple models training:

python run-autoANN.py -input_filepath models_config.xlsx -modIdStart 1 -modIdEnd 5

• Custom save directory:

python run-autoANN.py -input_filepath models_config.xlsx -modIdStart 1 -modIdEnd 5 -saveDir my_models

Input File (input-autoANN.xlsx) Configuration

Required Columns in Excel Input File

Column	Data Type	Description	Example
model_id	Integer	Unique identifier for each model	1, 2, 3...
database_path	String	Path to training database with the _database directory	"db-mpea-ys-ac-t25-seed1.xlsx"
x	String	Feature names separated by semicolons as defined in MAPAL	"(asymmetry,r_cov);(comp_avg,VEC);H_el"
y	String	Target variable name	"YS_MPa"
cols_to_keep	String	Columns in input database file to retain in output (comma-separated)	"alloy_name,phases,phase_code"
layer_units	String	Neural network layer sizes (comma-separated)	"64,32,16"
activation_functions	String	Activation functions (comma-separated)	"relu,relu,sigmoid"
optimizer	String	Optimizer name	"Adam"
learning_rate	Float	Learning rate value	0.001
loss_function	String	Loss function name	"mae"
Kfold_split	Integer	Number of cross-validation folds	5
n_runs	Integer	Number of training runs with random data-shuffling & model initialization in each run	3
patience_val	Integer	Early stopping patience	50
max_epochs	Integer	Maximum training epochs	1000
check_error_value	Float	Error convergence threshold	200
check_after_iterations	Integer	Check convergence after N epochs	100
tensorboard_logs	Integer	Enable TensorBoard logging (1 or 'true')	1

Output Structure

For each trained model, the following directory structure is created:

Directory/File	Description
M-{model_id}/	Main model directory
├── M-{model_id}_inputParam.xlsx	Input parameters used for training
├── M-{model_id}_xmin-feats.csv	Feature minimum values for normalization
├── M-{model_id}_xmax-feats.csv	Feature maximum values for normalization
├── M-{model_id}_CV-predictions-all-runs.xlsx	Cross-validation predictions
├── M-{model_id}_CV-performance-all-runs.xlsx	Performance metrics summary
├── M-{model_id}_model-checkpoints/	Training checkpoints directory
├── M-{model_id}_model-save-final/	Final trained models directory
├── M-{model_id}_csv-logs/	Training logs in CSV format
└── M-{model_id}_tensorboard-logs/	TensorBoard logs (if enabled)

Footnotes

1. Department of Metallurgical and Materials Engineering, Indian Institute of Technology Ropar, Rupnagar 140001, Punjab, India ↩ ↩² ↩³ ↩⁴