Machine Learning Approach to Simulation of Continuous Seeded Crystallization of Gibbsite

Continuous seeded crystallization is characterized by oscillations of particle size distribution (PSD) and liquor productivity. To describe these oscillations using analytical methods is a complicated task due to non-linearity and slow response of the process. This paper uses a statistical approach to the preparation of initial data, determination of the significant factors and arrangement of the said factors by their impact on the dynamics of crystal population development. Various methods of machine learning were analyzed to develop a model capable of forecasting the time series of particle size distribution and composition of the final solution. The paper proposes to use deep learning methods for predicting the distribution of crystals by grades and liquor productivity. Such approach has never been used for these purposes before. The study shows that models based on long short-term memory (LSTM) cells provide for better accuracy with less trainable parameters as compared with other multilayer neural networks. Training of the models and the assessment of their quality are performed using the historical data collected in the hydrate crystallization area at the operating alumina refinery