Design of ethylene oxide production process based on adaptive design of experiments and Bayesian optimization

Abstract

In process design, the values of design variables X for equipment and operating conditions should be appropriately selected for entire processes, including all unit operations, such as reactors and distillation columns, to consider effects between unit operations. However, as the number of X increases, many more simulations are required to search for the optimal X values. Furthermore, multiple objective variables Y, such as yields, make the optimization problem difficult. We propose a process design method based on adaptive design of experiments and Bayesian optimization. Selection of X values that satisfy the target values of multiple Y variables are searched, and simulations for the selected X values are then repeated. Therefore, the X will be selected by a small number of simulations. We verify the effectiveness of this method by simulating an ethylene oxide production plant.

Design of Molecules with Low Hole and Electron Reorganization Energy Using DFT Calculations and Bayesian Optimization

“…PTR is the probability that predicted y values will fall within a target range. Considering a normal distribution in which y values and their variance predicted by GPR models are the mean and variance, respectively, we can integrate the normal distribution from the lower limit of y , Y LOWER , to the upper limit, Y UPPER , and calculate PTR.…”

“…To solve this problem, Bayesian optimization (BO), , which is based on Gaussian process regression (GPR) , and uses not only predicted y values but also their standard deviations to find x candidates for the next experiments, was proposed. Based on predicted y values and their standard deviations, acquisition functions such as the probability of improvement (PI), expected improvement, mutual information, and probability in target range (PTR) are calculated, and the x candidates with the highest values of the functions are selected. BO allows us to properly search not only for interpolation of existing datasets but also for extrapolation regions, thus reducing the possibility of falling into local optimal solutions.…”

Integration of Materials and Process Informatics: Metal Oxide and Process Design for CO2 Reduction

“…In Bayesian optimization (BO), an acquisition function was calculated after the GPR models had predicted the y values, whose variances were also calculated. Because y has target ranges in this study, we used probability in the target range (PTR) 16 expressed in terms of probability to unify them into an acquisition function. PTR is the probability of the predicted y values falling within a target range of y .…”

“…Manufacturing the desired product via the RWGS-CL reaction process requires metal oxide and process designs and testing in pilot and actual plants. In recent years, both metal oxide 15 and process designs 16 have been conducted using machine learning. A statistical model y = f ( x ) is constructed using a dataset between x and y , which represent the synthesis conditions and the properties and activities, respectively, and the dataset is an experimental dataset in metal oxide design.…”

Process-Informatics-Assisted Preparation of Lithium Titanate Crystals with Various Sizes and Morphologies

“…BO is a method for calculating an acquisition function (AF) that considers the estimated Y values and the variation of the estimated Y values and for selecting a candidate of variables X with the highest value of an AF. , In BO, an estimated Y value and its variance are calculated with a GPR model for a new sample, and they are used to calculate an AF. In this study, we used the probability of improvement (PI) and probability in target range, which can be reasonably extended to multiple Y values, as an AF.…”