Machine Learning-Based Operational Modeling of an Electrochemical Reactor: Handling Data Variability and Improving Empirical Models

Abstract

Electrochemical reduction of carbon dioxide (CO2) has received increasing attention with the recent rise in awareness of climate change and the increase in electricity supply from clean energy sources. However, because of the complexity of its reaction mechanism and the largely unknown electron transfer pathways, the development of a first-principles-based operational model of a CO2 electrocatalytic reactor is still in its infancy. This work proposes a methodology to develop a feed-forward neural network (FNN) model to capture the input–output relationship of an experimental electrochemical reactor from experimental data that are obtained from easy-to-implement sensors. This FNN model is computationally efficient and can be used in real-time to determine energy-optimal reactor operating conditions. To further account for the uncertainty of the experimental data, the maximum likelihood estimation (MLE) method is adopted to construct a statistical neural network, which is demonstrated to be able to address a usual overfitting problem that occurs in the standard FNN model. In addition, by comparing the neural network with an empirical first-principles-based model, it is demonstrated that the neural network model achieves improved prediction accuracy with respect to experimentally determined input–output operating conditions. Finally, the insights obtained from the FNN model and the limitations identified of the empirical, first-principles model (EFP model) are used to propose specific modifications to the EFP model to improve its prediction capability.

Leave No Photon Behind: Artificial Intelligence in Multiscale Physics of Photocatalyst and Photoreactor Design

“…One possibility is to use supervised ML as a powerful tool for making a variety of predictions for catalyst and reactor properties, [ 84 , 85 , 86 , 87 , 88 , 89 , 90 ] where relevant features in catalyst and reactor design (i.e., electronegativity, band center, surface area, reaction conditions, and reactor geometry/topology) can serve as input features for training a model on desired outputs (i.e., product formation rate, selectivity, stability, and efficiency). [ 91 , 92 , 93 ] Within photocatalysis, ML has been successfully employed, for example, to predict perovskite materials (using features including from electronegativity, light intensity, photocatalyst quantity, and calcination temperature) [ 94 ] and layered double hydroxides (using elemental and structural features generated from external packages [ 95 , 96 ] and based on local chemical hardness) for water splitting, organic heterojunction photocatalysts (using electronic descriptors such as electron affinity and reorganization energy) for hydrogen production, [ 97 ] and optimal reaction conditions (flow rate and reactor temperature) for the degradation of dyes. [ 98 ] In addition, models have also been built based on literature data for establishing trends and connections for products, catalysts, and material properties for CO 2 photocatalysis in both the gas and liquid phases.…”

Review of Machine Learning for Hydrodynamics, Transport, and Reactions in Multiphase Flows and Reactors

“…In the literature, ML has played a vital role in chemical process systems engineering (PSE) design, optimization, and control, which was well reviewed and extensively studied by many researchers. − In this section, we will mainly focus on how to apply ML techniques to combine with the field data of flow and transport for multiphase device performance optimization. This aspect has received relatively little attention, compared with the topic in PSE.…”

Smart manufacturing inspired approach to research, development, and scale-up of electrified chemical manufacturing systems

“…In the electrochemical CO 2 reduction, FNN was used to model the steady-state production of thirteen product species and oxygenate selectivity. 94 Given the large number of product species and reactor variability, an FNN was chosen for this model because the steady-state performance of the reactor is assumed to be independent of time. The inputs to the FNN are the applied potential and rotation speed since these are independent variables that quantify the kinetic and mass-transfer rates, respectively.…”

“…The FNN model is described in more detail by Luo et al. 94 Figure 6 B shows the concentration predictions of ethanol, one of the most valuable products in this electrochemical reactor, and Figure 6 C shows the formic acid model without an overfit in the operational range described. 89 Additionally, this model can be used for management-level control and optimization.…”