Developing an online composition prediction for an HI-I-HO system using deep neural network

Tanaka, Nobuyuki ; Takegami, Hiroaki ; Noguchi, Hiroki  ; Kamiji, Yu; Myagmarjav, O.  ; Ono, Masato ; Sugimoto, Chihiro 

We developed a deep neural network method to predict the composition of the iodine-sulfur process of thermochemical water-splitting hydrogen production using measurable properties. Unlike conventional titration analysis, this approach allows a quick understanding of fluid composition, providing essential information for controlling operating conditions. This study focused on the HI-I-HO three-component system within the IS process. Using Gibbs phase rule, the DNN model was constructed using online measurable parameters, such as temperature, pressure, and density, as input conditions. The model was trained with experimental data, and the structural parameters were tuned. Composition prediction using actual trend data demonstrated good correlation with titration analysis measurements. Furthermore, the local interpretable model-agnostic explanations method was incorporated to gain insights into the significance of input parameters for compositions from the DNN model, providing valuable information on crucial parameters for effective composition control.