Accurately representing ice clouds in radiative transfer models is extremely challenging due to the high diversity of the shapes and habits of ice crystals found in these clouds. In addition, the impact of cloud conditions on microphysical processes and resulting crystal morphologies cannot be studied without having reliable measurements on the relative proportion of crystal habits inside a cloud. Although airborne optical array probes have existed for decades to address these issues, our ability to extract meaningful information out of the images produced by these probes has been limited because we were missing automatic, unbiased and reliable classification tools. Many attempts have been made, starting in the 1980s (Rahman et al. (1981) and Duroure (1982)) with feature-based decision trees. At the time, image recognition was only starting to take shape as a field of computer science and computational power was still a limiting factor. With the recent uprising of computer vision (Krizhevsky et al. (2017)), it is now possible to reproduce the human ability to identify complex objects without creating huge sets of parameters around specific data sets and trying to minimize inter-class variability and maximize intra-class variabilities, as has been recently documented in Praz et al. (2018). However, we believe that this leads inevitably to large bias and uncertainties when applied to actual data. In the presented study, a methodology for automatic ice crystal recognition using innovative machine learning was developed for the PIP instrument (and is currently under development for the 2DS). 6 classes have been defined specifically for the PIP to account for the 3 ice crystal formation processes (vapor deposition, riming and aggregation) for particles of max diameter above 2mm. More than 3000 images were used along with some data augmentation to provide a diverse and solid database that includes various types of aggregates found in Mesoscale Convective Systems (MCS) which will be a major focus in our future applications of this classification tool. In contrast, to the work of Xiao et al. (2019) we included classes such as fragile aggregates or rimed aggregates with high intra-class diversity in their shapes. Convolutional Neural Network (CNN) has been chosen as the method to achieve the best results together with the use of finely tuned dropout layers which guarantee higher quality in the classification results by creating multiple confirmation paths for a single habit. The network has been tested through random inspections on actual data, part of which was then assimilated in order to improve intra-class variability and reduce confusion. It showed accuracy above 93% on an independent test set with most of its confusions being explained by the assumed porosity/transition between certain classes for instance between rimed aggregates and fragile aggregates. We believe once this tool is ready for other airborne optical array probes, it will be able to foster an improved insight in ice clouds microphysical processes.