Datasets from "A framework to predict binary liquidus by combining machine learning and CALPHAD assessments"

Description

If you use this dataset, please cite the following publication where you can find more details: G. Deffrennes, K. Terayama, T. Abe, E. Ogamino, R. Tamura, A framework to predict binary liquidus by combining machine learning and CALPHAD assessments, Materials & Design (2023) https://doi.org/10.1016/j.matdes.2023.112111 In this work, 3 machine learning models are developed to predict the liquidus in binary systems. Model 1 predicts the formation of liquid miscibility gaps. Model 2 predicts the equilibrium onset temperature of solidification. Model 3 predicts the critical temperature of liquid miscibility gaps. The datasets that are used to build the models and evaluate their performance are found in the “CALPHAD Datasets” folder. They were collected from 466 CALPHAD assessments of binary phase diagrams listed in “Reference_list.csv”. It is emphasized that CALPHAD calculations in areas without experimental data were included in the training datasets. As a result, the datasets contain some unreliable data. This is especially true for the miscibility gap temperature regression dataset for Model 3, because (1) experimental data on liquid miscibility gaps are often lacking, especially at high temperatures, and (2) it includes extrapolations in metastable domains. The performance of the models was evaluated from a nested group cross-validation approach. The predictions generated during this step are found the “Predictions inside the datasets” folder. For Model 2 and Model 3, visual comparisons between the predictions and the CALPHAD assessments are provided. 1563 binary liquidus are predicted from the models (all the two-by-two combinations between the 64 elements missing from the datasets), and the results are found in the “Predictions outside the datasets” folder. They are obtained from the following procedure. Each model was trained on its whole dataset (expect for Model 3), after hyperparameters were tuned from a 10-fold cross validation. For Model 3, data on the B-Ga system that are unreliable were not taken into account. Besides, Model 3 is not valid in completely miscible systems, and was only used in systems where the probability of having a stable liquid miscibility gap is predicted by Model 1 to be greater or equal to 0.25. It is emphasized that Model 3 only gives rough estimates, and predictions for elements not included in the training dataset that comprises 32 systems and 29 elements should be taken with caution.

Steps to reproduce

This work can be reproduced from the information given in the associated research article, and from the “Algorithms_and_hyperparameters.pdf” file

Categories

Funding

Licence