Castro Garcia Abraham, Ching Phoebe Lim, So Richard Hy, Cheng Shuo, Boonyubol Sasipa, Cross Jeffrey S
Department of Transdisciplinary Science and Engineering, School of Environment and Society, Tokyo Institute of Technology, 2-12-1 S6-10, Ookayama, Meguro-ku, Tokyo 152-8552, Japan.
Bioengineering Graduate Program, Chemical and Biological Engineering Department, Hong Kong University of Science and Technology, 999077, Hong Kong.
ACS Omega. 2023 Oct 4;8(41):38148-38159. doi: 10.1021/acsomega.3c04275. eCollection 2023 Oct 17.
Both the conversion of lignocellulosic biomass to bio-oil (BO) and the upgrading of BO have been the targets of many studies. Due to the large diversity and discontinuity seen in terms of reaction conditions, catalysts, solvents, and feedstock properties that have been used, a comparison across different publications is difficult. In this study, machine learning modeling is used for the prediction of final higher heating value (HHV) and ΔHHV for the conversion of lignocellulosic feedstocks to BO, and BO upgrading. The models achieved coefficient of determination () scores ranging from 0.77 to 0.86, and the SHapley Additive exPlanations (SHAP) values were used to obtain model explainability, revealing that only a few experimental parameters are largely responsible for the outcome of the experiments. In particular, process temperature and reaction time were overwhelmingly responsible for the majority of the predictions, for both final HHV and ΔHHV. Elemental composition of the starting feedstock or BO dictated the upper possible HHV value obtained after the experiment, which is in line with what is known from previous methodologies for calculating HHV for fuels. Solvent used, initial moisture concentration in BO, and catalyst active phase showed low predicting power, within the context of the data set used. The results of this study highlight experimental conditions and variables that could be candidates for the creation of minimum reporting guidelines for future studies in such a way that machine learning can be fully harnessed.
将木质纤维素生物质转化为生物油（BO）以及对生物油进行升级一直是众多研究的目标。由于在反应条件、催化剂、溶剂和所使用的原料特性方面存在很大的多样性和不连续性，因此很难对不同出版物进行比较。在本研究中，机器学习建模用于预测木质纤维素原料转化为生物油以及生物油升级过程中的最终高热值（HHV）和ΔHHV。这些模型的决定系数（）得分在0.77至0.86之间，并且使用SHapley加性解释（SHAP）值来获得模型的可解释性，结果表明只有少数实验参数对实验结果有很大影响。特别是，对于最终HHV和ΔHHV而言，过程温度和反应时间在大多数预测中起主要作用。起始原料或生物油的元素组成决定了实验后获得的可能的最高HHV值，这与先前计算燃料HHV的方法一致。在所使用的数据集范围内，所使用的溶剂、生物油中的初始水分浓度和催化剂活性相显示出较低的预测能力。本研究结果突出了一些实验条件和变量，这些条件和变量有望成为制定最低报告指南的候选因素，以便能够充分利用机器学习。
