Toward Faster Operational Optimization of Cascaded MSMPR Crystallizers Using Multiobjective Support Vector Regression

Abstract

Mixed-suspension mixed-product removal (MSMPR) crystallization process is critical for optimal separation and purification operations in pharmaceutical and fine chemical industries. To achieve this, detailed mathematical model-based optimization is the current practice, which is reported as an extremely time-consuming exercise, prohibiting their online implementation. To facilitate faster optimization, a novel datadriven modeling and optimization algorithm has been proposed in this work. Limited number of high-fidelity data from the computationally expensive model were utilized to build surrogates using support vector regression (SVR). Inputs of MSMPR are assigned different, instead of same, kernel parameters, and multiple kernels were explored to capture the complex dynamics of the crystallization process. Loaded with a sample size estimation algorithm to reduce the computational load on model execution time, the complete formulation is solved using an evolutionary algorithm enabling evolution of optimal SVR-based surrogate models. From the Pareto optimal set of several such models, two instances with differing prediction error were selected and the optimization of MSMPR was performed. The study indicates 3 orders of magnitude faster optimization with the surrogate models and 87.54% savings in the number of expensive function evaluations compared to the physics-based MSMPR model facilitating online optimization of such a process.

A data driven approach in less expensive robust transmitting coverage and power optimization

“…Notably, the quality of surrogate construction strongly depends on the distribution of training points, sample size, and optimally adjustments of hyperparameters of the surrogate. Hyperparameter optimization along with control on a number of expensive simulations keeping in mind the model doesn't get overfitted has been performed in literature 59 . Even though stochastic learning theory is responsible for the development of Kriging, the bounds on the hyperparameters are typically fixed arbitrarily 57 , 60 .…”

Interpretation of Drop Size Predictions from a Random Forest Model Using Local Interpretable Model-Agnostic Explanations (LIME) in a Rotating Disc Contactor

“…These include the following: (i) robust optimization of computationally expensive networks employ rigorous cross-validation methodology for hyperparameter tuning 34 and (ii) multiobjective optimization of cascaded mixed-suspension mixed-product removal crystallizers. 35 In the current work, the data set is randomly split into a training set and a testing set using an 80− 20 split. We use R-squared (R 2 ) and Mean Absolute Error (MAE) as evaluation metrics to assess the performance of our model on the test set (Table 1).…”

Comparative study of artificial intelligence based multi‐modelling approach and optimization of photoreactor

“…The details of this method can be read in the article published by Chowdhury et al [8] The automation of meta parameter estimation by neural architectural search method can be found in other articles also. [10][11][12][13][14]…”

“…By applying neural architectural search (NAS), the correct model structure can be found balancing the accuracy of the prediction and overfitting issues of the model. [10][11][12][13][14] However, once the activation function is finalized, then the activation functions are not changed during neural model building. Only the best set of weights and biases are found to minimize the error.…”