Data for: Reiss et al. (2026) - Data-driven parameterization of Arctic Ocean turbulent mixing

Turbulent mixing is hypothesized to become increasingly important in a changing Arctic Ocean, influencing regional dynamics and global overturning. Arctic turbulence remains poorly characterized due to sparse observations, limitations of conventional parameterizations, and incomplete understanding. We train gradient-boosted decision trees on global microstructure turbulence data (>5000 profiles; ∼3×105 data points) to predict turbulent dissipation from more widely available hydrography. Global training is essential for predictive skill across non-Arctic and Arctic regions (R2=0.75, 75% within a factor of 2 of observations). Applied to pan-Arctic hydrographic observations, predictions show enhanced dissipation along margins and shelves and weaker values in basin interiors, consistent with current understanding. Applied to high-resolution model output, predictions reveal turbulent diffusivities consistent with observations and basin-integrated irreversible buoyancy flux, yet markedly different from the common K-Profile Parameterization - with implications for simulated tracer transport. Our machine learning parameterization can help assess climate-mixing feedbacks in projections while physics-based parameterizations mature. This dataset includes the compiled and preprocessed microstructure and hydrographic data and code for feature generation, machine learning model training, and turbulence predictions from hydrographic observations and LLC4320 output. It also includes subsets of the LLC4320 simulation: the hydrographic fields used for feature generation, the KPP diffusivity fields, and the resulting ML predictions. Notebooks are documented individually