Maestría en Analítica Aplicada · 2026

Exploration of machine learning interpretability methods for explainable chemical engineering models

Resumen

Chemical industrial operations like amine-based post-combustion carbon capture (PCC) exhibit complex, non-linear behaviors that challenge traditional data-driven modeling approaches. While simple models lack predictive power, advanced machine learning (ML) models like XGBoost can accurately capture these dynamics but often function as "black boxes", hindering their adoption due to a lack of transparency required for operational trust and decision-making in industrial environments. This thesis addresses this critical gap by investigating how model-agnostic interpretability techniques can unlock transparent, reliable, and actionable insights from high-performance ML models applied to industrial PCC processes. To achieve this, a high-fidelity XGBoost model predicting CO2 Capture Efficiency was developed and rigorously validated using Monte Carlo Cross-Validation, demonstrating significantly superior performance compared to a Linear Regression baseline by effectively modeling process non-linearities. Then, a comprehensive suite of model-agnostic interpretability methods was systematically applied to the validated XGBoost model. The results demonstrate that these techniques successfully elucidate complex relationships, revealing non-linear feature effects and interaction patterns missed by the linear baseline. Feature importance rankings differed significantly beyond the primary drivers, and methods like PDP/ALE visualized clear non-linear dependencies crucial for understanding process behavior. Comparative analysis highlighted how these advanced interpretations provide a more nuanced and accurate picture than simpler models.