Profile-aware prediction and resource-constrained optimization of malachite green photodegradation using Pd-BiFeO₃
Year: 2026
Type: conference-paper
Abstract: Machine-learning models for photocatalysis can appear overly accurate when time points from the same experiment enter training and test subsets. A 2 wt% Pd-BiFeO₃/malachite-green dataset was aggregated to 200 profile-time observations and evaluated using row-wise and profile-aware five-fold validation. CatBoost reached R² = 0.983 on the row-wise hold-out, whereas support-vector regression gave the strongest grouped generalization (mean R² = 0.865). Discrete constrained screening indicated that increasing catalyst loading from 1.0 to 1.5 g L⁻¹ reduced the predicted time to 90% degradation from 242 to 210 min and nominal irradiation-energy input by 13.4%. Profile-aware validation is essential for credible optimization.
DOI: 10.5937/ioc26229n
Institutions: University of Belgrade, RWTH Aachen University
Field: Energy
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W7221801641