Enantiodiscrimination
with single-molecule and single-shot resolution
is fundamental for the understanding of the fate and behavior of two
enantiomers in chemical reactions, biological activity, and the function
of drugs. However, molecular decoherence gives rise to spectral broadening
and random errors, offering major problems for most chiroptical methods
in arriving at single-shot-single-molecule resolution. Here, we introduce
a machine-learning strategy to solve these problems. Specifically,
we focus on the task of single-shot measurement of single-molecule
chirality based on enantioselective ac Stark spectroscopy. We find
that, in the large-decoherence region, where the ac Stark spectroscopy
without machine learning fails to distinguish molecular chirality,
in contrast, the machine-learning-assisted strategy still holds a
high correct rate of up to about 90%. Beyond this overwhelming superiority,
the machine-learning strategy also has considerable robustness against
variation of the decoherence rates between the training and testing
sets.