Modeling of Adsorption Equilibria through Gaussian Process Regression of Data in Dubinin’s Representation: Application to Water/Zeolite Li-LSX

Abstract

A thermodynamically consistent modeling of adsorption equilibrium data is essential in the modeling of adsorption heat pump cycles. Especially when experimental data are scarce, the approach taken by Dubinin and co-workers to map all adsorption data onto a single “characteristic curve” can be very useful. Different choices for obtaining this curve for modeling purposes are discussed and analyzed here using adsorption equilibrium data of water on zeolite Li-LSX as an example. It is shown that the fit function proposed by Dubinin and Astakhov results in a poor fit to the data. The use of arbitrary fit functions, however, leads to an overfitting of the data, as is exemplified through a cross-validation analysis. Gaussian process regression (GPR) is discussed as an alternative way to compute the “characteristic curve” and is shown to result in a good fit to the data while, by design, avoiding overfitting.

Zeolite NaY-Copper Composites Produced by Sintering Processes for Adsorption Heat Transformation—Technology, Structure and Performance

“…Furthermore, the pressure increase due to the inevitable leakage will bias the results to a lower adsorption potential. For this reason, we used only the gravimetric measurement data for the gaussian process regression (GPR) [60,61]. This results in a look-up table for W(A), which is used for calculating the equilibrium loading for a given set of temperatures and pressures.…”

Novel Adsorption Cycle for High-Efficiency Adsorption Heat Pumps and Chillers: Modeling and Simulation Results

“…The determination of this heat capacity and of the adsorption enthalpy q st (x i , T i ) is done as described in [22]. The characteristic curve required in the methodology of Dubinin [23] and used within this paper is fitted according to Gaussian process regression according to [24] and given in [10] (pp. 190-191).…”

“…190-191). This fitting method of Dubinin's characteristic curve leads to a good fit to measured adsorption equilibrium data while it avoids overfitting of these data that may occur with some alternative fitting methods [24]. The selected hyperparameters of the Gaussian process are given in Table 1.…”

“…Table 1. Hyperparameters for the fit of equilibrium loading of the working pair water-zeolite Li-Y using Gaussian process regression [24].…”