Honey Powder ML Optimisation
A full ML pipeline that reproduces and extends a published food-science paper (Maharani et al., 2024) on pilot-scale honey powder formulation. Replaced the original commercial RSM / Central Composite Design workflow with Gaussian Process Regression, Gradient Boosting, and XGBoost, then applied Pareto-front analysis and Bayesian Optimisation over just 30 experiments. Gradient Boosting and GPR beat RSM on held-out data, and the design space revealed 1,149 Pareto-optimal formulations the original study collapsed into a single number. Includes SHAP interpretability and an honest write-up of where XGBoost underperforms at small n.
Technologies
Python, Scikit-learn, XGBoost, Gaussian Process Regression, Bayesian Optimisation, SHAP, Jupyter
Links