Machine Learning for Predicting Hydrogen Yield in Microwave Pyrolysis: Engineering Practice of Interdisciplinary Research

This post introduces the open-source project MAP-Hydrogen-Yield-ML maintained by Roshni S.K. (GitHub link: https://github.com/RoshniSK9/MAP-Hydorgen-Yield-ML, released on June 16, 2026). Combining chemical engineering and machine learning, the project integrates 205 experimental data points from 13 studies to build a predictive model for optimizing the microwave-assisted pyrolysis (MAP) hydrogen production process. The project uses the XGBoost model to achieve optimal prediction performance and enhances model interpretability through SHAP analysis, providing data-driven guidance for process optimization in the clean energy transition.

In the global energy transition, hydrogen energy is one of the key technical paths to achieve carbon neutrality goals. Compared with traditional steam methane reforming for hydrogen production, hydrogen production via pyrolysis of biomass and waste is not only renewable but also solves solid waste disposal problems. As an emerging hydrogen production technology, microwave-assisted pyrolysis (MAP) is efficient but involves multiple complex factors such as raw material properties and operating parameters. Traditional trial-and-error experimental methods are inefficient. Machine learning can effectively address this challenge by guiding experimental design and process optimization through data-driven models.

The project extracts 205 experimental data points from 13 peer-reviewed research papers, covering various raw material types such as biomass, plastic waste, and municipal solid waste. Feature engineering follows a four-layer architecture: raw material characteristics (10 features, e.g., particle size, carbon content), microwave operation parameters (6 features, e.g., pyrolysis temperature, microwave power), microwave absorber characteristics (5 features, e.g., absorber type, dielectric constant), and catalyst characteristics (6 features, e.g., catalyst type, specific surface area). Missing value handling distinguishes between "MISSING_NA" (feature not applicable) and "MISSING_NR" (not reported in original studies) to ensure the accuracy of subsequent model training.

The project evaluates 6 machine learning models: tree ensemble models (XGBoost, Random Forest, Histogram-based Gradient Boosting Regression) and traditional models (Support Vector Regression, Ridge Regression, PCA+Linear Regression). Among them, the XGBoost model performs best with an R² of 0.76 on the test set. Its advantages include: automatic capture of high-order interaction effects between features, insensitivity to feature scaling, built-in feature importance evaluation mechanism, and robustness to outliers.

The project applies SHAP (SHapley Additive exPlanations) analysis to explain the prediction logic of the XGBoost model. Through visualization tools such as Beeswarm plots, Bar plots, Waterfall plots, and Dependence plots, it answers key questions: which features have the greatest impact, how feature values affect the prediction direction, and whether there are synergistic/antagonistic effects between features. The analysis results can directly guide experimental design, such as prioritizing the optimization of pyrolysis temperature and catalyst metal loading.

The project code adopts a modular design, with core logic encapsulated in the H2_pred_ML_models package, including modules like data.py (data loading and cleaning), preprocess.py (preprocessing pipeline), and models.py (model definition). Two environment configuration methods are provided: conda environment (environment.yml) and minimal dependencies (requirements.txt). A fixed random seed (random_state=30) ensures result reproducibility. Jupyter Notebook serves as a user-friendly interface, balancing ease of use and flexibility.

The current project has limitations: small data scale (205 samples), data heterogeneity (differences in experimental conditions across 13 studies), and incomplete features (e.g., microwave power distribution uniformity not included). Future improvement directions include: active learning to guide the selection of next-round experimental parameters, multi-task learning to predict both hydrogen yield and other product distributions, and physics-informed neural networks combining the advantages of pyrolysis kinetics equations and data-driven approaches.

This project demonstrates a typical paradigm for applying machine learning in the field of chemical engineering: problem definition → data integration → feature engineering → model selection → interpretability analysis → knowledge translation. It provides a complete reference template for researchers in materials science, chemical engineering, and other fields. The project not only offers a practical prediction tool but also transforms data science methods into engineering practice guidance through systematic data integration and interpretability analysis, contributing to the optimization of clean energy processes under carbon neutrality goals.