Accelerated exploration of multinary systems using high-throughput experiments and machine learning

To cite this work: Coupling mixture designs, high-throughput experiments and machine learning for accelerated exploration of multinary systems This repo is my sandbox to play with data from the Gitlab original repo that will stay mainly unchanged because it is linked to a scientific paper.

Repository of codes for experiments scheduling and machine learning modeling of experimental data. More details are given in README of each modules and in Documentation. The dataset used is stored in a separate dedicated cloud storage fitted for big data and can be consulted and downloaded from here.

This repo contains the high-definition figures as they should have appeared in lossless png in the article before being butchered by the editor in jpg despite our claims.

Part 1: Paper codes (Python and Matlab)

This part describes the exact codes and methods used for the paper.

Experiments Planification

This Matlab GUI interface allows to prepare the experimental work. It gives a set of 1D linear gradients or 2D planar gradients for 3 to 7 element systems. It can minimize the total number of samples or the experiments price.

Requirements: Matlab2019 or more

PyTerK

These Python modules and notebooks perform iterative k-fold crossvalidation for Scikit-learn or Keras models, on any datasets, with parallelization of works. Settings are written in yml files, trainings are performed in "run" notebooks and results are visualized in "report" notebooks.

Requirements:

• Install following libraries, via pip or conda
  • tensorflow - keras
  • pandas
  • scikit-learn
• Create environment variables:
  • path to folder that contains the datasets DATASETS_DIR=$path/to/datasets/
  • path to folder that will contain the training results : RUN_DIR=$path/to/run/

Multiple Regression

These Python modules and notebooks perform iterative k-fold crossvalidation for statsmodel multilinear regression, on any datasets. Training and report are performed in one notebook.

Requirements:

• Install following libraries, via pip or conda
  • statsmodels.formula.api
  • pandas
  • scikit-learn
• Create environment variables in bash_profile:
  • path to folder that contains the datasets: export DATASETS_DIR = $path/to/datasets/
  • path to folder that will contain the training results : export RUN_DIR= $path/to/run/ NB: for windows, see this help

Evaluations Models Tables

These notebooks are a complement to report in PyTerK. They allowed to gather evaluation metrics/scores in tables and export them as csv, while notebooks allows to visualize them.

Requirement: same as PyTerK

Galanov Model

This module allows to compute elastic-plastic zone $\frac{b_s}{c}$, the constrain factor $C$ and ductility characteristic $\delta_H$ by solving equations proposed by Galanov et al (Galanov, Ivanov, et Kartuzov, Improved Core Model of the Indentation for the Experimental Determination of Mechanical Properties of Elastic-Plastic Materials and Its Application.)

Requirements:

• Install following libraries, via pip or conda
  • pandas
• Create environment variables:
  • path to folder that will contain the training results : DATASETS_DIR=$path/to/datasets/ NB: for windows, see this help

Pareto Front

This module allows to detect Pareto optimal composition, for two antagonistic criteria/properties.

Requirements:

• Install following libraries, via pip or conda
  • pandas
• Create environment variables:
  • path to folder that will contain the training results : DATASETS_DIR=$path/to/datasets/ NB: for windows, see this help

Plot Predictions

These Matlab modules plot Machine Learning prediction. It can scatter composition point of plot alpha shape of compositions points. User can add color or point size to plot along the predicted properties of the composition points.

Requirements:

• Matlab2019 or later
• Create environment variables:
  • path to folder that will contain the training results : DATASETS_DIR=$path/to/datasets/ NB: for windows, see this help

Part 2: Full Matlab implementation

Authors: Adrien GUILLE and Marek BRAUN, Grenoble Institute of Technology, France

This part describes a full Matlab implemenation using the Statistics and Machine Learning Toolbox.

Just run Main.m and enjoy the full process in one click !

Requirements:

• Matlab2019 or later