Overview

GEOS-FP (Goddard Earth Observing System, Forward Processing) is a data product created from GEOS, an integrated Earth system model. GEOS-FP uses a numerical weather prediction model — physical and chemical — to provide real-time meteorological data, including atmospheric concentration of several pollutants at 2.5μm: dust, sea salt, organic carbon, brown carbon, and sulfate.

These predicted concentrations do not match observed concentrations from ground stations and on-site monitoring. We used a combined convolutional neural network and long short-term memory model to correct those biases.

GEOS-FP provides several other parameters that may improve the model, such as atmospheric boundary layer height, wind speed and direction, and land-based organic matter measurements. We analyzed the correcting power of each.

Training

1. Basic dataset manipulation (NumPy, Pandas)
2. Understanding of common data formats used in atmospheric sciences (HDF5 and NetCDF3/NetCDF4)
3. General understanding of neural networks, specifically convolutional and time series based
4. General atmospheric and earth science principles were a bonus, taught along the way

Group meeting once a week (mandatory), with optional meetings with professors and scientists one to three times a week.