Biochar-Augmented Anaerobic Digestion System: Insights from an Interpretable Stacking Ensemble Deep Learning.

Area of Science:

• Environmental Science and Engineering
• Biotechnology
• Artificial Intelligence

Background:

• Anaerobic digestion (AD) is crucial for waste management and renewable energy production.
• Biochar augmentation can enhance AD efficiency, but process optimization remains challenging.
• Predictive modeling is needed to optimize complex AD systems with biochar.

Purpose of the Study:

• To develop an interpretable deep learning model for optimizing biochar-augmented anaerobic digestion (AD).
• To predict methane yield and ensure process stability under various conditions.
• To provide a practical framework for real-world AD optimization.

Main Methods:

• Compiled extensive experimental data on feedstock, operational conditions, biochar properties, and stability indicators.
• Developed a stacking ensemble deep learning model integrating CNNs and LSTMs.
• Employed hyperparameter tuning, permutation importance, and SHAP for model optimization and interpretability.
• Implemented a model-based global optimization framework and a graphical user interface.

Main Results:

• The stacking ensemble model achieved high internal predictive accuracy (mean R² 0.91-0.94) and strong generalization (R² 0.68 on external data).
• The model outperformed individual deep learning models in predicting methane yield.
• Interpretability analysis identified critical factors influencing methane production and AD stability.
• The optimization framework successfully tailored conditions for high methane yield and stability.

Conclusions:

• The interpretable deep learning model provides a robust framework for optimizing biochar-augmented AD systems.
• This approach enhances methane yield prediction and ensures process stability for efficient waste-to-energy conversion.
• The developed tools facilitate practical implementation, advancing resource recovery and sustainable waste management.