Aerosol-cloud interactions contribute significant uncertainty to modern climate model predictions. Analysis of complex observed aerosol-cloud parameter relationships is a crucial piece of reducing this uncertainty. Here, we apply two machine learning methods to explore variability in in-situ observations from the NASA ACTIVATE mission. These observations consist of flights over the Western North Atlantic Ocean, providing a large repository of data including aerosol, meteorological, and microphysical conditions in and out of clouds. We investigate this dataset using principal component analysis (PCA), a linear dimensionality reduction technique, and an autoencoder, a deep learning non-linear dimensionality reduction technique. We find that we can reduce the dimensionality of the parameter space by more than a factor of 2 and verify that the deep learning method outperforms a PCA baseline by two orders of magnitude. Analysis in the low dimensional space of both these techniques reveals two consistent physically interpretable regimes—a low pollution regime and an in-cloud regime. Through this work, we show that unsupervised machine learning techniques can learn useful information from in-situ atmospheric observations and provide interpretable results of low-dimensional variability.