Facies interpretation from wire-line logs has traditionally been performed through comparison of core-observed facies distributions and associated log response, from which log-based, deterministic algorithms are developed for facies prediction in wells lacking core control. In contrast, stochastic unsupervised learning analyzes and automatically clusters recurring well log data associations without calibration to core observations. Using petrophysical and core description data collected from the Late Duvernay Formation of Alberta, Canada, this study investigates whether unsupervised machine learning detects lithologic attributes at higher resolution than the traditional deterministic approach. The unsupervised machine learning methodology nonnegative matrix factorization with k-means clustering (NMFk) is applied to identify recurring petrophysical groups independent of core observations. The NMFk groups are compared to four depositional facies associations independently predicted through deterministic assignment of well log cutoffs established by comparing log response to core-observed facies association distributions. Depositional facies associations include the undifferentiated Ireton Shale that overlies the Duvernay and the open to transitional basin, restricted basin, and allochthonous basinal carbonates of the Duvernay. Four NMFk groups are identified; three groups coincide with varying shale lithologies and one group with carbonate lithologies. All of the groups are similar to the depositional facies associations recognized by the deterministic approach. In addition, NMFk partitions the deterministic restricted basin facies association into subdivisions that are likely distinguished by unique petrophysical responses related to variations in organic richness and associated matrix pore fluid.