ML-hardness-MPEAs

Publication details:

Distilling physical origins of hardness in Multi-Principle-Element Alloys directly from ensemble neural networks

• Authors : Dishant Beniwal ¹, P. Singh ², S. Gupta ², M.J. Kramer ², D.D. Johnson ²³, Pratik K. Ray ¹
• Journal : npj Computational Materials
• DOI : https://doi.org/10.1038/s41524-022-00842-3

Contents

This repository contains codes and database used for creating neural network models for prediction of hardness in multi-principal element alloys.

File "db_HEAs.csv":

Database file - 218 HEAs with their hardness, phase information and normalized feature values.

File "run_ANN_HV_prediction.py":

Creates ANN model using parameters defined in "Input_ANN.txt" file. A new directory is created to store trained models and training/validation results.

Requirements|recommended versions: Python|3.8.1 ; pandas|1.0.3 ; numpy|1.18.2 ; scikit-learn|0.22.2 ; tensorflow|2.2.0rc2 ; keras|2.3.1

File "Input_ANN.txt":

Input file to define all the model parameters. Make modifications only after "[" symbol and don't use spaces while modifying parameters. Refer to original file for reference.

• "project_name" - A directory with this name will be created. All results will be stored here.

• "database" - Write database filename here.

• "y" - Column name in database that will be used as target (keep this as 'VHN' as hardness is being predicted).

• "x" - Name of features to be included in model (these can be changed freely, but don't add space between feature names), feature added here must be present as a column in "db_HEAs.csv" file.

• "layer_units" - No. of units in sequential hidden layers; This also controls no. of hidden layers; last layer is output layer (1 unit, don't change last layer).

• "activation_functions" - Activation function for sequential hidden layers; last layer is output layer (relu, don't change last layer).

• "loss_function" - Loss function used for error quantification (we used 'mean absolute error' - mae).

• "optimizer" - Optimizer used for loss minimization (we used Adam).

• "learning_rate" - Learning rate in backward propagation (we used contant lr=0.02).

• "iterations" - No. of iterations for which model will run (we used 5000).

• "save_after_iterations" - No. of iterations after which current model will be saved (we used 100 i.e. model state is saved after every 100 iterations).

• "check_error" - Threshold error that must be attained after (check_after_iterations); otherwise model will re-initialize all parameters and will start from beginnning. This ensures that model will either converge or will restart.

• "check_after_iterations" - No. of iterations after which threshold error will be checked.

File "f_extract_input_data.py":

Contains "f_extract_input" function that is used in main script "run_ANN_HV_prediction.py" for extracting parameters defined in "Input_ANN.txt" file.

File "f_ANN.py":

Contains "f_ANN_model" function that is used in main script "run_ANN_HV_prediction.py" for creating the neural network model using parameters extracted from "Input_ANN.txt" file.

Footnotes

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

2. Ames Laboratory, US Department of Energy, Ames, IA 50011, USA ↩ ↩² ↩³ ↩⁴

3. Materials Science & Engineering, Iowa State University, Ames, IA 50011, USA ↩