Effect sizes calculated using mean difference for burnt-unburnt study designs and mean change for before-after desings. Outliers, as defined in the methods section of the paper, were excluded prior to calculating effect sizes.

This repository contains the datasets produced from the characterizations of the quinary Nb-Ti-Zr-Cr-Mo, and predictions made by Machine Learning models.

Experimental work

Gradients of composition were characterized by:

EDX for composition evaluation, with an error of 1% on atomic and mass composition

nanoindentation for the measurement of the elastic modulus (E) and hardness (H)

EBSD : from each map we extract the Confidence Index CI and Image Quality IQ that are indicator of crystallinity. CI is also used to define phase classes (0 for amorphous, 1 for crystalline)

XRD: from each diffractogram we extract a phase class (0 for amorphous, 1 for crystalline): raw data are available in XRD.zip

Different datasets are built:

Raw_data associate to each composition the EBSD CI, IQ, EBSD phase class, and the elastic modulus (E) and hardness (H) computed by the software TestWork Analysis without any correction. For each composition, 5 measurement replications were performed.

Raw_data_corrected contains the EBSD CI, IQ, EBSD phase class, and the 5 replications per compositions of E and H corrected through Oliver and Pharr model.

Compo_E_H_threshold correspond to Raw_data_corrected in which we have thresholded values of E and H. We removed all composition such that E < 10 GPa and all H < 2 GPa, as they correspond to nanoindentation test failures.

Compo_E_wo_outliers and Compo_H_wo_outliers: Dixon test allows to identify outliers on E replications and H replications, that are removed to give each dataset. Each composition is associated to replications of E or H that were not identified as outliers.

Averaged_data: each composition is associated to EBSD CI, IQ, EBSD phase class, and with average values of E and H replications without outliers.

Data_averaged_mechanical_model: add to previous data the other mechanical properties computed with Galanov model from E and H experimental results: relative characteristic size of the elastic-plastic zone under the indenter (x = \frac{b_s}{c}), the constrain factor (C) – linking yield strength and hardness – and the ductility characteristic (\delta_H) – ratio of plastic deformation and total deformation. It also contains (\frac{E²}{H}).

Database_XRD: each composition is associated to phase class defined from XRD diffractograms

The dataset_initial.zipl contains the experimental results with an initial 20-gradients sets which screen preferably the center of Nb-Ti-Zr-Cr-Mo. It contains all the kind of datasets.

The dataset_adding_binaries.zip contains the experimental results for the initial 20-gradients + additional binary gradients Nb-Ti binary 1), Nb-Cr (binary 2) and Cr-Mo (binary 3). It contains the data without outliers, averaged data and XRD database.

Predictions of Machine Learning Models from experimental datasets

Machine Learning models are trained to predict properties from compositions: Random Forest (RF), Support Vector Machine (SVM) and Neural Network (NN) models.

Model assessment (i.e. choosing best hyper-parameters for each model) was performed on Compo_E_wo_outliers for E prediction, Compo_H_wo_outliers for H prediction, and on Averaged_data and Database_XRD for phase prediction. Results of model trainings are given in ModelAssessment.tar.gz.

The best model of RF, NN and SVM are trained on all datasets: results are given in Train_model_xx.tar.gz. Training the same model with datasets with more or less outliers for E and H predictions allows to see the effect of outliers on the results.

The best models of RF and NN are then trained adding iteratively the binaries: results are in tarball Train_model_xx_adding_binaries.tar.gz

These tarball are to be used with PyTerK modules available here.

The models then predict, for all atomic compositions of Nb-Ti-Zr-Cr-Mo, with 2%at steps, the associated properties:

predictions_XX contain atomic compositions associated to predicted CI, IQ, EBSD phase class, XRD phase class, E, H, for each kind of model.

Predictions_XX_mechanical_model contain the same data with other mechanical properties computed with Galanov model from E and H predictions: relative characteristic size of the elastic-plastic zone under the indenter (x = \frac{b_s}{c}), the constrain factor (C) – linking yield strength and hardness – and the ductility characteristic (\delta_H) – ratio of plastic deformation and total deformation. It also contains (\frac{E²}{H}).

The prediction_initial.zip contains the predictions made for all the model families with initial datasets.

The predictions_adding_binaries.zip the predictions made with the best model (determined with the initial dataset) trained with the initial dataset+ binaries