Interpretable machine learning for predicting and evaluating hydrogen production via supercritical water gasification of biomass

Abstract Supercritical water gasification (SCWG) of biomass for hydrogen production is a clean and promising technology. However, due to many factors involved in SCWG process, including biomass properties and process parameters, it is time consuming and capital intensive to evaluate the multi-dimensional relationship between them, as well as the hydrogen production capability using the experimental method. Therefore, it is necessary to develop an accurate model to predict and evaluate SCWG process in an economic way. This study established four machine learning models to predict hydrogen production via SCWG of biomass, interpreted the inner workings of the optimal model and evaluated the performance of SCWG. The results suggested that random forest (RF) model outperformed gaussian process regression, artificial neural network and support vector machine models for predicting H2 yield (R2 = 0.9782). Feature importance and partial dependence analysis combined with the RF model were used to visually present the relative importance and average partial relationship selected biomass properties and process parameters for H2 yield. The contour plots based on the RF model indicated that maximum hydrogen reaction efficiency (45.6%) and exergy efficiency (43.3%) were achieved when biomass with a high O content and low H/C ratio were used as feedstock. This study demonstrated that the machine learning model is a practical tool for predicting hydrogen production. Meanwhile, the interpretation of the model can clarify the influence of the variables involved in SCWG on hydrogen production, which is helpful for selecting the appropriate feedstock and optimize the process parameters in the experiment or practical engineering.