Predicting Bio-oil Yield from Biomass Pyrolysis Using Machine Learning-based Tools

Biomass pyrolysis involves the thermal decomposition of biomass constituents to yield valuable compounds to be exploited as biofuels and/or platform chemicals. Modelling biomass pyrolysis is challenging, but modern AI methods may give a valuable contribution to prediction of process yields, provided high-quality datasets are available. The published literature on the subject mostly refers to limited datasets, usually only a few hundred records, which are inadequate for robust AI applications. This work presents a dataset of about 500 observations with no missing values, compiled from published data on bio-oil/bio-liquid production via fixed-bed pyrolysis of different biomass. The dataset includes physicochemical properties of the biomass, key pyrolysis operating conditions, and bio-liquid yield. Each observation was carefully standardized to resolve inconsistencies in the terminology and/or lack of standardization. Best results, obtained from XGBoost, showed a MAE of 2.0 and an R2 of 0.8. Critical analysis of results demonstrates that AI applied to biomass pyrolysis data displays very good predictive ability. However, the ability to reproduce known relationships among key variables of the biomass and of the process appears to be more problematic. This is well shown by analysis of PDP plots, where some inconsistencies with known trends emerge when assessing the influence of selected variables on bio-liquid yield. Moreover, pronounced discrepancies with previously published studies by other research groups are observed when analyzing directional trends.