Here a 3D convolutional recurrent neural network performance to forecast PM <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2.5</inf> and O <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</inf> at the different stages of the lockdown in Colombia are presented. The network makes a 24 hours forecast and is evaluated using 4 statistical parameters that are then averaged for every lockdown period to quantify the change in the model precision due to lockdown policies. For Colombia, PM <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2.5</inf> predictions are better for partial and full lockdown, due to the high values produced by biomass burning (pre-lockdown) that are sub-estimated by the network. This does not happen for O <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</inf> where the model is fairly similar for the three periods, but in the full lockdown, the model has a slightly worst performance, probably because O <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</inf> in full-lockdown have an abnormal growth that leads to a sub estimation of its magnitude, even though the behavior is well represented by the model.On the other hand, the model is more precise for PM <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2.5</inf> in the full and partial lockdowns possibly due to the biomass burning in pre-lockdown that produces abnormal increases of the pollutant. In terms of the ozone, the pre and partial lockdown have better performance since in the full lockdown the O <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</inf> rises to higher values than usual, which has not happened before, so the model was not able to learn this behavior in a precise manner. The topography and the pressure levels seem to be the principal source of uncertainty for the model (apart from the biomass burning).