Casting hybrid twin: physics-based reduced order models enriched with data-driven models enabling the highest accuracy in real-time

Abstract

Knowing the thermo-mechanical history of a part during its processing is essential to master the final properties of the product. During forming processes, several parameters can affect it. The development of a surrogate model makes it possible to access history in real time without having to resort to a numerical simulation. We restrict ourselves in this study to the cooling phase of the casting process. The thermal problem has been formulated taking into account the metal as well as the mould. Physical constants such as latent heat, conductivities and heat transfer coefficients has been kept constant. The problem has been parametrized by the coolant temperatures in five different cooling channels. To establish the offline model, multiple simulations are performed based on well-chosen combinations of parameters. The space-time solution of the thermal problem has been solved parametrically. In this work we propose a strategy based on the solution decomposition in space, time, and parameter modes. By applying a machine learning strategy, one should be able to produce modes of the parametric space for new sets of parameters. The machine learning strategy uses either random forest or polynomial fitting regressors. The reconstruction of the thermal solution can then be done using those modes obtained from the parametric space, with the same spatial and temporal basis previously established. This rationale is further extended to establish a model for the ignored part of the physics, in order to describe experimental measures. We present a strategy that makes it possible to calculate this ignorance using the same spatio-temporal basis obtained during the implementation of the numerical model, enabling the efficient construction of processing hybrid twins.