A framework to predict binary liquidus by combining machine learning and CALPHAD assessments - Archive ouverte HAL Accéder directement au contenu
Article Dans Une Revue Materials & Design Année : 2023

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

Résumé

Knowledge of the liquidus is important for the design and processing of many materials. For instance, deep eutectics are important for the design of metallic glasses, and recently multi-principal element alloys have been designed based on eutectic compositions or melting temperatures extrapolated from binary liquidus data. In this study, we provide a general framework for predicting binary liquidus only from the properties of the pure elements and thermodynamic properties calculated by Miedema’s model. Our framework combines three machine learning models that are trained and evaluated on liquidus data collected from 466 CALPHAD assessments of binary phase diagrams. The first model predicts the formation of liquid miscibility gaps with a prediction accuracy of 95.3%, outperforming the empirical Mott model. The second and third models predict the equilibrium onset temperature of solidification and the critical temperature of liquid miscibility gaps, respectively. An important feature of our models is that they can give indications of the presence of congruent melting phases and eutectics. Using our framework, we predict the liquidus in 1563 binary systems not included in our CALPHAD dataset, many of which are unknown. By collecting more data, our framework will continue to grow towards better liquidus prediction.
Fichier principal
Vignette du fichier
2023Deffrennes_Liquidus.pdf (2.04 Mo) Télécharger le fichier
Origine : Publication financée par une institution
Licence : CC BY - Paternité

Dates et versions

hal-04187514 , version 1 (24-08-2023)

Licence

Paternité

Identifiants

Citer

Guillaume Deffrennes, Kei Terayama, Taichi Abe, Etsuko Ogamino, Ryo Tamura. A framework to predict binary liquidus by combining machine learning and CALPHAD assessments. Materials & Design, 2023, 232, pp.112111. ⟨10.1016/j.matdes.2023.112111⟩. ⟨hal-04187514⟩
29 Consultations
43 Téléchargements

Altmetric

Partager

Gmail Facebook X LinkedIn More