Supplement for "On using neural networks to predict mean Age-of-Air from long-lived tracers", submitted to ACP

These datasets contain data used for the work to be published in ACP with the running title "On using neural networks to predict mean Age-of-Air from long-lived tracers".

Each dataset contains random samples from CLaMS simulations between 420 K and 900 K potential temperature. The full data can be requested from the authors, see publication. Read the filenames as follows:

The keyword after "data" (pure, lat, time, all, theta) indicate the variables used as features for the neural networks. "Pure" contains only the 6 chemical species CH4, F11, F12, F22, N2O and SF6. "Lat" contains the latitude in addition to these chemical species. "Time" contains a time variable in addition to the chemical species. "Theta" contains the potential temperature in addition to the chemical species. "All" contains all aforementioned variables.

The XXXX number indicates the year of the simulation from which the data was sampled.

Each dataset contains 50000 samples from the first of every month of the indicated year of the simulation. These samples are split into "training" data (40000 samples) and "testing" data (10000 samples).

In this work the data was scaled to range (0, 1) using Min-Max-Scaling. The scale is chosen such that all values in the selected potential temperature range is scaled accordingly. The parameters used for this scaling is included in each dataset as the variable PARAMETERS, which includes the minimum value, maximum value, mean value and standard deviation for each variable and month. Naturally the data will lie inside the (0, 1) range, but will not have hard limits at 0 and 1.