This study investigates an adaptive ensemble machine learning approach to predict HTC reactor output parameters under severe data scarcity conditions. Three complementary learning algorithms were selected: Neural Networks (NN), capable of capturing complex non-linear relationships but sensitive to data scarcity; Support Vector Regression (SVR), known for its robustness in high-dimensional spaces and generalization capability with limited samples; and TreeBagger (Random Forest), which provides stable predictions through ensemble averaging of decision trees. These models were integrated through a hybrid strategy that dynamically combines model predictions using optimized weights determined through systematic grid search. Results demonstrate significant improvements in critical parameters. Where combined predictions do not improve upon the best single model, the approach automatically defaults to the superior individual predictor, ensuring robust performance across all parameters. This adaptive framework proves particularly valuable for HTC process optimization under data scarcity, offering a practical solution that maximizes information extraction from limited experimental data.
Machine Learning Ensemble Strategy for HTC Reactor Modelling Under Data Scarcity
Bartolomeo Cosenza;
2025-01-01
Abstract
This study investigates an adaptive ensemble machine learning approach to predict HTC reactor output parameters under severe data scarcity conditions. Three complementary learning algorithms were selected: Neural Networks (NN), capable of capturing complex non-linear relationships but sensitive to data scarcity; Support Vector Regression (SVR), known for its robustness in high-dimensional spaces and generalization capability with limited samples; and TreeBagger (Random Forest), which provides stable predictions through ensemble averaging of decision trees. These models were integrated through a hybrid strategy that dynamically combines model predictions using optimized weights determined through systematic grid search. Results demonstrate significant improvements in critical parameters. Where combined predictions do not improve upon the best single model, the approach automatically defaults to the superior individual predictor, ensuring robust performance across all parameters. This adaptive framework proves particularly valuable for HTC process optimization under data scarcity, offering a practical solution that maximizes information extraction from limited experimental data.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
