Otimização do rendimento da etapa de precipitação para aumento de produtividade do Processo Bayer

Abstract

The Bayer process, the main route for alumina production, has a potential bottleneck for increasing overall productivity in its precipitation step. Optimizing the yield of this phase is crucial to improving the efficiency and economics of the process. The present work optimizes the yield of the precipitation step of the Bayer process through the development of an innovative hybrid model. This approach combines the fundamentals of phenomenological modeling, which describes the physicochemical mechanisms underlying the process, with the predictive capability of models based on machine learning techniques to identify optimized operational conditions. This way, it is possible to overcome the limitations of purely phenomenological models, which may have difficulties in dealing with the complexity and nonlinearities inherent in industrial processes. The integration with machine learning techniques allows for the exploration of large volumes of operational data and the identification of complex patterns, enabling the prediction of precipitation yield as a function of various process variables. To achieve the proposed objective, the adopted methodology first involves the elaboration of a detailed phenomenological model of the precipitation step of the Bayer process, incorporating the main phenomena of nucleation, growth, and agglomeration of aluminum hydroxide particles. For the model to be more robust and realistic, the dissociation of chemical species and pH were considered. Subsequently, a machine learning algorithm was implemented in Python, using synthetic data generated from the phenomenological model. A comparison of different hybrid modeling architectures was performed: series, parallel, and series-parallel. Through the comparative analysis of the hybrid modeling architectures, it is observed which one presents the best performance in predicting the precipitation yield. The implementation of the hybrid model results in an increase in productivity, demonstrating the potential of integrating phenomenological modeling and machine learning for the optimization of complex chemical processes.