Deep learning for kinetics parameters identification: A novel approach for multi-variate optimization
Antonello Raponi, Daniele Marchisio
Abstract
This work presents an innovative deep-learning approach for multi-variate optimization, focusing on the identification of Mg(OH)2 precipitation kinetics parameters. The study employs three distinct experimental datasets, one for the Population Balance Model (PBM) fitting and two for validation. These datasets explore the impact on Particle Size Distributions (PSDs) of (i) increasing the initial reactant concentrations from 0.125 to 1 M and (ii) decreasing the flow rate from 12 to 4 m/s, both in a T-mixer, (iii) increasing the initial reactant concentration over a wider concentration range from 0.01 to 1 M in a more complex Y-mixer system. Leveraging PBM, we create a dataset to train a Neural Network (NN), referred to as the ‘mirror model,’ which predicts kinetics parameters based on experimental sizes. Notably, the PBM, fitted with dataset (i), excels at describing changes in flow rate (dataset (ii)) and substantial reductions in reactant concentrations in the Y-mixer (dataset (iii)), even though these conditions were not encountered during the fitting step. Key Performance Indicators (KPIs) reveal that the mirror model consistently outperforms two widely used algorithms, Conjugate Gradient (CG) and Particle Swarm Optimization (PSO), highlighting its remarkable potential for practical applications.