A Machine Learning-Based Multi-Objective Framework for Optimizing Fruit-Peel Waste- Derived Bio-Pesticide Formulations

Abstract: Fruit-processing waste can be a cheap source of biologically active plant compounds, but turning it into a workable bio-pesticide depends on choosing the right composition and operating conditions. This study presents a prediction-and-optimization workflow for 200 fruit-peel formulations, in which regression models estimate Mortality % and Repellency %. Five models were compared: Random Forest, Gradient Boosting, XGBoost, Support Vector Regression (SVR), and Extra Trees. In the saved five-model evaluation, which used the available five-predictor CSV, XGBoost gave the best held-out test results for Mortality % (R² = 0.854, MAE = 3.049, RMSE = 3.621), and Gradient Boosting did best for Repellency % (R² = 0.773, MAE = 3.631, RMSE = 4.795). These figures still need to be reconciled with the seven-predictor feature set described in the original manuscript before final submission. The earlier Pareto analysis kept 19 non-dominated formulations and picked FPF_177 as a computational screening candidate, though biological confirmation is still needed.

Keywords: fruit-peel waste; bio-pesticide; machine learning; XGBoost; Support Vector Regression; Extra Trees; Pareto optimization.