MASS-CONSERVING PHYSICS-INFORMED AUGMENTATION AND FOURIER FEATURE NETWORKS FOR SMALL-DATA PREDICTION OF MOLYBDENITE (MO₂S) LEACHING KINETICS

Abstract

Molybdenum remains a strategic metal for advanced steels and catalysis, while environmental and energy pressures are accelerating interest in hydrometallurgical leaching routes for molybdenite (MoS₂). Predicting leaching kinetics is difficult because the process is highly nonlinear and strongly influenced by reagent chemistry and gas–liquid conditions, yet experimental datasets in metallurgical laboratories are often extremely small. This manuscript develops a hybrid, data-efficient machine-learning approach designed specifically for small-data settings. The method combines physics-informed data augmentation that enforces strict mass conservation with a Fourier Feature Network intended to reduce spectral bias and better capture sharp kinetic transitions. Using only six experimental measurements, the resulting model achieves high predictive accuracy on held-out data (R² = 0.9793, MAE = 1.61%) and maintains stable generalization without evidence of train–test divergence. The study concludes that physically admissible augmentation coupled with Fourier-enriched representations can produce reliable kinetic surrogates from minimal data, supporting in-silico screening and optimization of leaching conditions for process design and control.