Development of a stacked machine learning model to compute the capability of ZnO-based sensors for hydrogen detection

Abstract

Zinc oxide (ZnO) nanocomposite sensors decorated with various dopants are popular tools for detecting even low hydrogen (H₂) concentrations. The nanocomposite's chemistry, temperature, and H₂ concentration impact the success of hydrogen sensors. Extensive laboratory-scale studies were conducted to investigate the effect of these variables on sensor performance, there is currently no model to relate the nanocomposite's sensitivity to its influential variables. This study proposes a stacked model by integrating Extra tree and XGBoost (eXtreme Gradient Boosting) regressor to precisely relate the sensitivity of the ZnO-based sensor to the nanocomposite's chemistry, H₂ concentration, and temperature. The model's accuracy is superior to that of conventional artificial neural networks, achieving outstanding prediction results with mean absolute error (MAE) = 0.11, mean squared error (MSE) = 0.31, mean absolute percentage error (MAPE) = 1.14%, and R-squared (R²) = 0.9994 based on 208 actual sensor sensitivities. Also, the designed stacked model predicts 206 experimental records with relative error ranges from −4% to 8%. Applicability domain analysis confirms the validity of almost all experimental measurements (200 out of 208 records). Trend and relevancy analyses indicated that the sensor sensitivity intensifies with increasing hydrogen concentration and decreasing temperature. The reduced graphene oxide (rGO) dose initially improves sensor sensitivity toward hydrogen detection up to a maximum value and then continuously decreases it. The analysis of variance approves that the ZnO-Co₃O₄ sensor has the maximum value of least squares average = 42.3 for hydrogen detection over its experimental conditions. This study provides valuable insights for designing efficient ZnO-based sensors for hydrogen detection, ultimately contributing to safe hydrogen transportation/utilization.