All life forms operate metabolism in constant flux. Metabolic fluxes offer a direct readout of cellular state, detailing the rates and driving forces of metabolic pathways. However, indirect, iterative solvers for mapping isotope patterns from tracing experiments onto metabolic fluxes leave much of cellular state uncharted. Here, we streamline metabolic flux quantitation by innovating a machine learning framework, ML-Flux, that deciphers complex isotope labeling patterns. We train neural networks using isotope pattern-flux pairs across central carbon metabolism from 26 key 13C-glucose, 2H-glucose, and 13C-glutamine tracers. ML-Flux takes variable-size isotope labeling patterns as input, imputes missing isotope patterns, and outputs mass-balanced metabolic fluxes. Computation of fluxes using ML-Flux is more accurate and faster than that of leading metabolic flux analysis software employing a least-squares method. Our biochemical networks and machine learning models constitute a curated and growing online knowledgebase of metabolic flux and free energy to democratize quantitative metabolic profiling.