Berlin 2024 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
MM: Fachverband Metall- und Materialphysik
MM 4: Data Driven Material Science: Big Data and Workflows I
MM 4.10: Vortrag
Montag, 18. März 2024, 12:45–13:00, C 243
Physics-informed neural network for predicting the Gibbs free energy — •Clement Paulson1, Amin Sakic2, Vedant Dave3, Elmar Rueckert3, Ronald Schnitzer1, and David Holec1 — 1CDL KnowDAS, Department of Materials Science, Montanuniversität Leoben, Austria — 2Department of Materials Science, Montanuniversität Leoben, Austria — 3CPS Lab, Montanuniversit Leoben, Austria
We employ a physics-informed neural network approach in conjunction with the CALPHAD formalism to determine the Gibbs free energy of alloys. The Gibbs free energy, essential for CALPHAD simulations, is determined by predicting the Redlich-Kister parameter using a composite neural network utilizing novel descriptors derived from the atomic, composition-based, and thermodynamic properties of elements. The composite neural network comprises a low-fidelity network trained on CALPHAD-generated mixing enthalpies and a high-fidelity network trained on experimental mixing enthalpies. These two models are further connected to a physics-informed neural network, which determines the Redlich-Kister parameters. The predicted Redlich-Kister parameters can then be directly implemented into a thermodynamic database file for immediate use with existing CALPHAD software. This approach holds promise for expediting materials development and phase stability determination. Comparative experimental results highlight the accuracy and potential of this deep learning-based method, offering a novel path for forecasting the Gibbs free energy in multi-component systems and accelerating the development of databases.
Keywords: physics informed neural network; CALPHAD; mixing enthalpy; gibbs energy; redlich-kister polynomial