Interpretable Deep Learning Empowers Biomass Pyrolysis: AI-Driven Precise Optimization of Renewable Energy

The InterpretableDNN_PyrolysisModel project, open-sourced by the Tang Laboratory at Wuhan University, combines interpretable AI (SHAP) with deep neural networks to provide high-precision predictions and transparent insights for the biomass pyrolysis process. This project aims to address the limitations of traditional experimental optimization methods and the "black box" problem of deep learning, bridging the gap between model decisions and chemical intuition, and facilitating the precise utilization and optimization of global renewable energy feedstocks.

Biomass pyrolysis is a key technology for producing renewable energy products such as bio-oil and biochar, involving complex physical and chemical reactions influenced by multiple factors like raw material properties and temperature. Traditional experimental methods are time-consuming and labor-intensive, making it difficult to cover the vast parameter space; while deep learning has strong predictive capabilities, its "black box" nature hinders the understanding of mechanisms. The growing global demand for renewable energy drives the development of precise prediction and optimization technologies.

Dual-Module Architecture: 1. Product Yield Prediction Module: Integrates industrial/elemental analysis of raw materials and temperature parameters to predict bio-oil, char, and gas yields, identifying key influencing factors via SHAP; 2. Kinetic Parameter Prediction Module: Predicts the activation energy of raw material degradation, validated with TG/DTG experimental data. Six-Stage Hyperparameter Optimization: Comprehensive tuning covering hidden layer architecture, learning rate, dataset division, momentum coefficient, activation function, and training strategy to ensure stable model performance.

The project is applied to the study of co-pyrolysis of municipal sludge with biomass such as rice husks and sawdust. Through integration of Monte Carlo simulation and DNN models, thousands of ratio combinations are explored to identify optimal strategies (pollution reduction + energy recovery enhancement). The model can predict changes in activation energy of mixed raw materials, determine the ratio that reduces reaction energy and improves biochar quality; the supporting script (missing_value_handler.py) handles missing values in heterogeneous sludge data.

The project code is well-organized: separate directories are set up for product yield/activation energy prediction, including training, evaluation, and visualization code; environment application code and datasets are organized separately. The hyperparameter optimization dataset, tuning logs, and related literature are open-sourced to support reproduction and expansion.

Scientific Value: Achieves the combination of deep learning's high predictive capability and understanding of physical and chemical mechanisms. Industrial Significance: Provides decision-making tools for pyrolysis process optimization, helping enterprises quickly evaluate raw material formulas, optimize reaction conditions, and reduce experimental costs. The interpretability feature makes AI prediction results trusted by engineers, which is key to AI's implementation in traditional industries.

InterpretableDNN_PyrolysisModel provides an innovative solution for the precise utilization of renewable energy through interdisciplinary integration (AI + bioenergy). The project has both academic research value and industrial practical value, and will play an important role in advancing global carbon neutrality goals.