Khan Muzammil, Surendra K C, Baniya Sachita, Rhymer Jay H, Khanal Samir Kumar
Department of Molecular Biosciences and Bioengineering (MBBE), University of Hawai'i at Ma̅noa, 1955 East-West Road, Honolulu, Hawaii 96822, United States.
Department of Civil, Environmental and Construction Engineering (CECE), University of Hawai'i at Ma̅noa, 2540 Dole Street, Honolulu, Hawaii 96822, United States.
Environ Sci Technol. 2025 Jul 29;59(29):15236-15250. doi: 10.1021/acs.est.5c05051. Epub 2025 Jul 18.
This study presents a comprehensive approach for optimizing biochar-augmented anaerobic digestion (AD) system through an interpretable stacking ensemble deep learning model. Extensive experimental data were compiled, incorporating feedstock characteristics, operational conditions, and biochar properties, alongside stability indicators such as pH, volatile fatty acid concentrations, alkalinity, and total ammonia nitrogen levels. The proposed model integrates different configurations of convolutional neural networks and long short-term memory networks within a stacking ensemble framework, effectively capturing complex interdependencies within the AD process and improving methane yield predictions. Optimized through advanced hyperparameter tuning, the model achieved high internal predictive accuracy (mean of 0.91-0.94 and root-mean-square error of 60.85 mL CH/g volatile solids) and demonstrated strong generalization with an of 0.68 on external independent lab-scale datasets, outperforming all individual models. Post-hoc interpretability analysis using permutation importance and Shapley Additive Explanations (SHAP) identified critical factors influencing methane production and stability indicators. A model-based global optimization framework was implemented to tailor optimal operational conditions for real-world scenarios, ensuring high methane yield while maintaining process stability. Additionally, a user-friendly graphical interface was developed to facilitate the practical implementation of the predictive model. This work provides a robust framework for optimizing the AD process with biochar augmentation, enhancing resource recovery and waste management.
本研究提出了一种综合方法，通过可解释的堆叠集成深度学习模型来优化生物炭强化厌氧消化（AD）系统。收集了大量实验数据，包括原料特性、运行条件和生物炭性质，以及诸如pH值、挥发性脂肪酸浓度、碱度和总氨氮水平等稳定性指标。所提出的模型在堆叠集成框架内整合了卷积神经网络和长短期记忆网络的不同配置，有效捕捉了AD过程中的复杂相互依赖关系，并改进了甲烷产量预测。通过先进的超参数调整进行优化后，该模型实现了较高的内部预测准确性（均值为0.91 - 0.94，均方根误差为60.85 mL CH/g挥发性固体），并在外部独立实验室规模数据集上以0.68的R2表现出很强的泛化能力，优于所有单个模型。使用排列重要性和Shapley加法解释（SHAP）进行的事后可解释性分析确定了影响甲烷产生和稳定性指标的关键因素。实施了基于模型的全局优化框架，为实际场景定制最佳运行条件，确保在保持过程稳定性的同时实现高甲烷产量。此外，还开发了一个用户友好的图形界面，以促进预测模型的实际应用。这项工作为利用生物炭强化优化AD过程、提高资源回收和废物管理提供了一个强大的框架。
