Phytoplankton is an essential component in the functioning of marine ecosystems and in the carbon cycle. It is therefore essential to assess its variability and its main drivers. However, unlike seasonal and interannual variations, fluctuations of phytoplanktonic biomass and communities on decadal to multi-decadal timescales remain hampered by the lack of long-term observations at global scale and the uncertainties related to the complex balance of the processes that control their fate. These processes are imperfectly and diversely parameterized in biogeochemical models, limiting their use to document long-term phytoplankton variability. Yet, it is crucial to detect natural low-frequency cycles in phytoplankton biomass (and thus carbon fluxes) because they can enhance, weaken or even mask climate-related trends. In this context, the inter/transdisciplinary DREAM project proposes to investigate and benchmark different deep learning (DL) frameworks (learned from satellite and in situ observations) to emulate past and future multi-decadal time-series of surface phytoplankton biomass and communities. This approach will allow us to assess the relative contribution of the different processes (e.g. physical, predation, community structures) involved in phytoplankton dynamics over the last decades in response to natural climate low-frequency variability but also to past and future anthropogenic forcing. Ultimately, DREAM will also contribute to characterizing and better constraining the uncertainties in the climate projections of the different Earth System Models gathered in the Coupled Model Intercomparison Project Phase 6 (CMIP6).