(130e) Digital Twin for Aluminum Smelting

Building digital twins for manufacturing processes like aluminum smelting using first principles modeling has been considered challenging owing to the scale and complexity of such processes. A considerable mismatch is normally observed between measured data and predicted values of desired variables. This mismatch is usually because of uncertainty in the information fed to the model such as model assumptions, parameters, correlations from literature, etc. A systematic approach to identify such missing information would be to pose the model as an input-reconstruction problem. While a typical input reconstruction problem is solved using optimization-based algorithms where the objective would be to minimize the deviation between data and prediction, it may not be suitable for complex processes such as aluminum smelting. In this work we present a novel method to reconstruct missing information using a combination of first principles modeling and data analysis. The input reconstruction problem is reformulated as a set of differential algebraic equations (DAE) using measured data. Solving these DAEs help identify the missing input information and other unknown states in one shot. This method gives a complete data set (reconstructed inputs and measured data) which is used to train machine learning (ML) algorithms to build new correlations and predict unknown inputs directly in the digital twin model thereby giving more accurate predictions of desired process parameters.