An interpretable ensemble machine learning model for predicting carbon dioxide adsorption on magnesium oxide-based...

Area of Science:

• Materials Science
• Chemical Engineering
• Data Science

Background:

• Accurate prediction of carbon dioxide (CO2) adsorption is vital for developing effective carbon capture technologies.
• Magnesium oxide (MgO)-based sorbents show promise for CO2 capture, but predictive models are needed to optimize their performance.

Purpose of the Study:

• To develop an interpretable machine learning framework for predicting the CO2 adsorption capacity of MgO-based sorbents.
• To identify key material and process parameters influencing adsorption capacity.
• To provide a tool for designing improved CO2 adsorbents.

Main Methods:

• Evaluated fourteen candidate machine learning models.
• Selected top models for Bayesian hyperparameter optimization using Optuna.
• Constructed a weighted ensemble model integrating entropy weight method and TOPSIS.
• Utilized SHapley Additive exPlanations (SHAP) for feature importance analysis.

Main Results:

• The optimized ensemble model achieved an R2 of 0.9903 on an independent test set.
• Material morphology, particularly granular, was the most significant feature (43.8% contribution).
• Process parameters like time (23.3%) and temperature (14.1%) were also major contributors.
• Optimal adsorption conditions identified as medium temperature (200-400°C) and adsorption time (100-250 min).

Conclusions:

• The developed machine learning framework provides accurate predictions for CO2 adsorption on MgO-based sorbents.
• Material morphology and process parameters significantly influence adsorption capacity.
• The study offers mechanistic insights for designing high-performance MgO-based CO2 adsorbents and a deployable web application for real-time predictions.