Designed nutrient-gradient phenotyping enables predictive ionomics of tropical leafy greens

Accurate prediction of edible-leaf mineral composition is challenging because tissue ionomes integrate nutrient supply, plant identity, growth dilution, water-balance regulation and cumulative environmental exposure. We combined a designed nutrient-gradient experimental framework with machine learning to predict leaf ionome using 1,163 Chinese spinach (Amaranthus dubius) and Chinese broccoli (Brassica oleracea Alboglabra Group) plants grown across 16 designed nutrient-gradient experiments in a semi-controlled tropical greenhouse. Targeted perturbations of N, P, K, Ca, Mg, Fe and Mo were combined with cumulative transpiration, integrated microclimate, shoot dry weight and laboratory quantification of 12 harvested-leaf mineral targets. The resulting ionome was broad, non-Gaussian and species dependent, with coordinated off-target shifts also evident for minerals whose supplied concentrations were held constant. We benchmarked AutoGluon and TabPFN using five-fold leakage-free, distribution-balanced grouped cross-validation, in which all biological replicates from each dosing condition were withheld together. Across previously unseen nutrient conditions, pooled out-of-fold R2 ranged from 0.72 to 0.92, with broadly comparable performance between architectures. Learning curves showed rapid early gains for several targets but persistent condition-level generalization gaps for total reduced nitrogen, K, Mn and Zn. Model-agnostic Shapley additive explanations identified target-specific combinations of nutrient inputs, plant type, substrate, cumulative environment and physiological traits, while cross-model attribution agreement varied among minerals. Finally, reduced-feature models were tested on an independent grower-operated cohort of 99 plants. Calibration was closest for NO3, Mg and B, whereas Mn and Na showed substantial external bias. These results establish leakage-aware predictive ionomics as a robust framework for post-harvest lab-free crop mineral estimation while defining the calibration and sensing requirements for commercial deployment.