Climate predictions and weather forecasting strongly rely on simulations of the Earth’s oceans and atmosphere turbulent dynamics. But the simulation of turbulent processes is so computationally expansive that it is only possible to resolve the largest physical scales. The representation of unresolved scales in these simulations is therefore a key source of un- certainty and its modeling is still an open problem. Recently, machine learning techniqueshave been receiving growing attention for the design of parametrizations and subgrid-scale models. In this thesis, we explore the impact of explicitly embedding law invariances in neural networks trained to represents the small-scale dynamics of a scalar quantity advected by a turbulent flow. We ...