A catalyst-aware explainable machine learning framework for biodiesel yield prediction over metal-doped biochar and...

Area of Science:

• Catalysis and Sustainable Chemistry
• Chemical Engineering and Process Optimization
• Machine Learning Applications in Chemical Processes

Background:

• Biodiesel production relies on complex interactions, making yield prediction difficult.
• Existing machine learning (ML) models lack catalyst-specific frameworks and struggle with small datasets and interpretability.
• Accurate prediction is crucial for catalyst screening and process optimization in sustainable biodiesel production.

Purpose of the Study:

• Develop an explainable ML framework for accurate biodiesel yield prediction.
• Integrate catalyst, feedstock, and operating condition descriptors for a unified model.
• Enhance understanding of factors influencing biodiesel yield for process optimization.

Main Methods:

• Utilized a literature-derived dataset of metal-doped biochar and activated carbon (AC) catalysts.
• Integrated catalyst, feedstock, and operating condition descriptors.
• Augmented sparse data via curve digitization and evaluated six ML models.
• Applied SHAP and CatBoost for explainability analysis.

Main Results:

• A neural network model achieved high predictive accuracy (RMSE: 3.27%, MAE: 1.64%, R ²: 0.95) on unseen data.
• Linear regression showed poor performance, confirming the system's nonlinearity.
• Explainability analysis identified alcohol-to-oil ratio, reaction time, temperature, catalyst amount, and feedstock acid value as key factors.

Conclusions:

• The proposed ML framework offers an accurate and interpretable approach for biodiesel yield prediction.
• The framework facilitates data-driven catalyst screening and process optimization for sustainable biodiesel production.
• Model generalization was confirmed through validation with an independent experimental dataset.