Modeling of biodiesel production using optimization designs from literature: aiming to reduce the laboratory workload

Abstract

This study explores non-linear tree-based learning algorithms for modeling biodiesel reactions. A dataset of 3038 reaction samples from 111 published studies was compiled, each optimizing distinct biodiesel reaction systems. Key operational parameters were selected to represent the dataset's diversity. Random forest (RF) and gradient boosting regressor (GBR) models were employed to predict biodiesel yield across the various reaction systems. GBR, with 1000 estimators and a tree depth of 5, achieved the best performance (R² = 0.744, RMSE = 10.783). The global GBR model was comprehensively evaluated for accuracy and physical relevance, with proposed applications in component screening and reaction optimization using the DIRECT-l (DIviding RECTangles - locally biased version) algorithm. Additionally, an experimental reaction was optimized via the global model and DIRECT-l, then refined using a retrained local model for improved system-specific predictions. These models offer researchers a data-driven approach to selecting and optimizing biodiesel reactions, reducing laboratory time and improving predictive accuracy for specific systems.