Abstract. Radiocarbon may serve as a powerful dating tool in palaeoceanography, but its accuracy is limited by the need to calibrate radiocarbon dates to calendar ages. A key problem is that marine radiocarbon dates must be corrected for past offsets from either the contemporary atmosphere (i.e. “reservoir age” offsets) or a modelled estimate of the global average surface ocean (i.e. delta-R offsets). This presents a challenge because the spatial distribution of reservoir ages and delta-R offsets can vary significantly, particularly over periods of major marine hydrographic and/or carbon cycle change such as the last deglaciation. Modern reservoir age and delta-R estimates therefore have limited applicability. While forward modelling of past R-age variability has been proposed as a means of resolving this problem, this requires accurate a priori knowledge of past global radiocarbon budget closure (i.e. production, and cycling), which we currently lack. In this context, the construction of empirical regional marine calibration curves could provide a way forward. However, the spatial reach of such calibrations and their robustness subject to (uncertain) temporal changes in climate and ocean circulation would need to be tested. Here, we use unsupervised machine learning techniques to define distinct regions of the surface ocean that exhibit coherent behaviour in terms of their radiocarbon age offsets from the contemporary atmosphere (R ages), regardless of the causes of R-age variability. We apply multiple algorithms (k-means, k-medoids, and hierarchical clustering) to outputs from two different numerical models spanning a range of climate states, forcings, and timescales of adjustment. Comparisons between the cluster assignments across model runs confirm some robust regional patterns that likely stem from constraints imposed by large-scale ocean and atmospheric physics. At the coarsest scale, regions of coherent R-age variability correspond to the major ocean basins. By further dividing basin-scale shape-based clusters into amplitude-based subclusters, we recover regional associations, such as increased high-latitude R ages, or the propagation of R-age anomalies from regions of deep mixing in the Southern Ocean to upwelling sites in the eastern equatorial Pacific, which cohere with known modern oceanographic processes. We show that the medoids (i.e. the most representative locations) for these regional sub-clusters provide significantly better approximations of simulated local R-age variability than constant offsets from the global surface average. This remains true when cluster assignments obtained from one model simulation are applied to simulated R-age time series from another. Further, model-based clusters are found to be broadly consistent with existing reservoir age reconstructions that span the last ∼30 kyr. We therefore propose that machine learning provides a promising approach to the problem of defining regions for which empirical marine radiocarbon calibration curves may eventually be generated.