Overcoming Sensitivity Barrier in 13C NMR Using Frequency–Time Dual-Domain GANs: Quantitative Reaction Monitoring at Natural Abundance

Low sensitivity has long limited the use of heteronuclei such as 13C and 15N NMR for quantitative chemical measurements, confining reaction monitoring largely to proton-detected experiments. Although highly sensitive, proton NMR often becomes unreliable due to severe signal overlap. Here we show that machine-learning-enabled heteronuclear NMR overcomes this limitation, enabling accurate, time-resolved reaction monitoring using natural-abundance 13C nuclei without isotopic enrichment or hardware modification. We introduce a frequency–time dual-domain generative adversarial network that enforces physical consistency between the free-induction decay and its frequency-domain spectrum. This dual-domain constraint suppresses noise while preserving peak intensities and lineshapes essential for quantitative analysis. We demonstrate that baseline-based signal-to-noise ratios become unreliable after ML-assisted denoising due to non-uniform noise floors, and therefore adopt a true signal-to-noise ratio based on peak-area variance. Using this metric, our approach achieves a two- to three-fold improvement in quantitative sensitivity, comparable to detection using cryogenically cooled probe-heads. Applied to the acid-catalysed inversion of sucrose, the method enables reliable 13C-based kinetic analysis with sub-minute acquisition times per spectrum, establishing a general framework for extending quantitative reaction monitoring beyond proton NMR.