Turbulent mixing across density surfaces (i.e., diapycnal mixing) in the ocean interior is key to sustaining the meridional overturning circulation and its global regulation of heat, carbon, and nutrient distributions, as well as other climatically and environmentally important tracers (Talley et al., 2016). Such turbulence is primarily excited at the ocean surface by winds, or at the bottom boundary via flow impingement on topography (Garabato & Meredith, 2022). The spatio-temporal variability of turbulence makes its measurement especially challenging. However, turbulence can leave an imprint on vertical temperature (T) and salinity (S) profiles obtained from hydrographic surveys. T, S, and depth (Z) are regularly sampled through global international programs, such as ship-based efforts like WOCE (Gouretski & Koltermann, 2004), GO-SHIP (GO-SHIP, 2018), GEOTRACES(GE-OTRACERS, 2019), or globally-distributed floats deployed by the Argo Program (Argo, 2000) (see Supporting Information S1 for a visual summary, and Davis et al. ( 2019) for a review (Davis et al., 2019)-hereafter we refer to Supplementary Materials as SM). While turbulence characteristics may be inferred from these T, S, Z data (Polzin et al., 2014;Whalen et al., 2012), such estimates involve many assumptions and uncertainties.The gold standard in measuring turbulence in the ocean interior is represented by ship-deployed microstructure profiler observations, which include concurrent sampling of T, S, and Z, but are limited in number due to their technical complexity and cost (Shroyer et al., 2018). In this study, we train machine-learning models on a unique collection of observations from microstructure field programs enabling prediction of turbulence characteristics based on T, S, Z, and topographic data, rendering our approach applicable to major global surveys that do not measure turbulence directly. Our aim is to demonstrate that such predictions from microstructure-trained physics-inspired machine-learning models yield better estimates for dynamically-significant quantities than classical finestructure parameterizations.