River deposits are rich in sedimentological patterns that determine their petrophysical properties and industrial value (e.g. hydrocarbon reservoirs and aquifers), as well as providing each deposit with individual characteristics. Great variability in characteristics exists in modern river systems; their behaviour is expected to further diverge under environmental and climate change scenarios, and much greater variation exists when we consider all of geological time. In view of this diversity in sedimentological patterns, a series of fundamental questions arise. Have we recognized all significant patterns? And, are all patterns that we recognize significant? These questions underpin both our ability to characterize river deposits and to interpret the singular dynamics that gave rise to individual deposits. Machine learning provides a means of re-analysing data to understand bias and uncertainty within our data domains, and can therefore be used to improve our understanding of ancient and modern rivers from existing datasets. Herein, we apply machine learning techniques to classify sediment core data taken from one of the world’s largest rivers: the Río Paraná, Argentina. We apply self-organizing maps for unsupervised clustering of sedimentary structures in the cores. The methods are suited to the analysis of different types of data, including core, borehole geophysics, and outcrops of varying size and quality. Early results illustrate that machine learning classification can: (1) recognize depositional units based on sedimentary structures and their location; and (2) simplify complex patterns of primary sedimentary structures, and present them as sedimentary facies with varying abundance, similarity and degrees of certainty. The success of these initial experiments opens up new opportunities for, for example: (i) automated distinction of depositional environments; (ii) the quantification of vertical and lateral trends in sedimentary deposits; and (iii) an improved interpretation of the generic and singular dynamics that give rise to river deposits. Thus, the results illustrate that machine-learning methods can identify significant patterns from sedimentary cores, including patterns that reveal areas which require further systematic research. Being able to classify sedimentological information in such a semi-automatic way opens new avenues for the analysis of much larger data domains, the interpretation of which would otherwise require vast manual expert effort. Automated detection of sedimentological features from data also enables the quantification of levels of uncertainty within the data domain. As such, machine learning provides a rigorous pathway to integrate uncertainty information into data-driven predictive models, and allows identification of ‘poorly understood’ aspects of river systems.