Introduction Ammonia has attracted attention as a potential hydrogen carrier for the realization of a hydrogen society. In the development of ammonia electrolysis catalysts operating under low-temperature and atmospheric pressure conditions, it is necessary to conveniently measure the amount of hydrogen generated by the competing proton reduction reaction as well as ammonia. In our previous study, we reported that by combining Deep Neural Networks (DNN) with genetic algorithms and Bayesian optimization to analyze data obtained from the rotating ring-disk electrode (RRDE) method, it is possible to simultaneously quantify hydrogen and ammonia with higher accuracy compared to using nonlinear least squares. In this study, we report on the improvement of the data acquisition method and the influence of the data normalization method on the prediction accuracy of ammonia and hydrogen partial pressure. Experimental RRDE measurements were performed in a phosphate buffer solution (pH 7) using a Pt ring-glassy carbon (GC) disk electrode (ring inner diameter 7 mm, outer diameter 9 mm, electrode area 25.13 mm2) as the working electrode, an Ag|AgCl electrode as the reference electrode, and a Pt electrode as the counter electrode. All potentials are quoted with respect to the reversible hydrogen electrode (RHE). Prior to measurements, the electrolyte was purged for 20 minutes with 100 mL/min H2-Ar gas at any desired partial pressure. The disk electrode was maintained at open circuit potential while cyclic voltammograms (CVs) of the Pt ring electrode were measured in the potential range of 1.4 to 0.15 V at a rotation speed of 600 rpm. The measurements were performed by altering the ammonium ion concentration (10−7 to 10−1 M) and hydrogen partial pressure (1 to 0 atm). To ensure a clean Pt ring surface, a potential of 1.5 V was applied for 1 minute before each measurement. PyTorch was used as the machine learning library. CVs were normalized using two methods: standardization with hydrogen oxidation current cut-off (P1) and standardization without hydrogen oxidation current cut-off (P2). DNN models were developed to predict the concentration of ammonium ions (10N M) or hydrogen partial pressure (H atm) using the standardized data. The NSGA-II algorithm was used to optimize the predictions by the DNNs. The nodes of the DNN were considered as explanatory variables, and 1−R2 served as the objective variable, where R2 is the coefficient of determination between the predicted and experimental values. Results and discussion The relationship between the predicted and experimental values of the ammonium ion concentration and hydrogen partial pressure using the nonlinear least-squares method (Broyden) and the DNN model with optimized nodes are shown in Fig. 1(a) and Fig. 1(b), respectively. The highest coefficients of determination (R2) in the plots of predicted versus measured values were obtained when P1 was used for ammonium ion concentration estimation (R2=0.96074) and when P2 was used for hydrogen partial pressure estimation (R2=0.99998). These values improved by 15.47% and 6.35%, respectively, compared to using the alternative normalization methods. These results demonstrate that prediction accuracy can be improved by tailoring the data normalization method for each predicted substance and employing multiple DNN models. Figure 1