Activity: Talk or presentation › Invited talk › Science to science
Description
During thermomechanical processing of titanium alloys in the β-domain, the β-phase undergoes dynamic recovery by forming new low-angle grain boundaries, which in turn form a new subgrain structure. This phenomenon is described by physics-based models that predict the microstructure evolution as a function of the material parameters and the deformation conditions (average strain, average strain rate and temperature). However, the complexity of these models in terms of rate equations and the mutual connection between their internal variables increases the computational time in FE simulations. In this work, two physics-informed Machine Learning (ML) metamodels are established to enable predicting the evolution of the substructure formation during hot deformation of a Ti-17 alloy with initial fully recrystallized microstructure and initial grain size of approximately 500 µm. To this end, two ML algorithms are used, namely decision tree regression (DTR) and Artificial Neural Networks (ANNs). The performance of the obtained metamodels with respect to the actual physical models is evaluated using the coefficient of determination (R²) and the root-mean-square error (RMSE). The evolution of subgrain size predicted by model and metamodel are compared during hot isothermal deformation at different temperatures and strain rates. Both metamodels can predict and generalize the target outputs properly (R2≥0.99) for unseen datasets. Despite the exceptional results for both metamodels, the physics-informed ANN is the most suitable to train large datasets, while DTR performs better for small datasets.
Period
13 Sept 2022
Event title
19th International Cconference on Metal Forming 2022