Trajectory Cluster Computations

This page describes the computational details of an approach used to examine if there is a long-term trend in trajectory climatology that can be shown using cluster analysis. The computational approach uses the HYSPLIT model to create the master trajectory database. The April 2025 Revision (v5.4.0) was used for the 48 h backward isobaric trajectory calculations starting at 612 locations over the globe (at 10 degree intervals), four times per day (0,6,12,18 h UTC), from 1500 m AGL. A detailed description of the HYSPLIT modeling system was published in BAMS.

The trajectory calculations were performed from 1950 through 2024 using meteorological data from the NCAR-NCEP global reanalysis. To facilitate the computations, the data were downloaded in a HYSPLIT compatible format from the NOAA-ARL data archive server. The gridded data fields are available every 6 h at 2.5 deg spatial resolution. Although there are better quality data sets available, they do not all afford the same ease of use.

A detailed discussion about how to configure HYSPLIT for various simulations can be found in the Users Guide and online tutorials. The 75 year computational period results in about 109,500 trajectories per location or 67,014,000 trajectories globally. The computational and data management issues are not addressed in the Guide or other online sources. Although all the computations could be done with the HYSPLIT package as downloaded, the time required on the single computer (3.1 GHz Quad-Core Intel i5) available for this task would not have been realistic. Several changes were introduced. The trajectory computations were run as a matrix, requiring only a single 48 h simulation and a single meteorological data file for all the global starting locations for each trajectory start time, rather than 612 simulations reading the same data file. This approach required a new program to deconstruct the single global trajectory endpoints file into an individual file for each location, the file structure required for the clustering program. The clustering program also needed to be modified to make it faster and eliminate its huge memory requirements, which then permitted the analysis of a much larger number of trajectories; 15,000 per decade.

The trajectory calculation can represent air mass movement. A cluster analysis for each trajectory origin location is used to determine the upwind flow patterns by decade: 1950-1959, 1960-1969, 1970-1979, 1980-1989, 1990-1999, 2000-2009, 2010-2019, and the incomplete period 2020-2024. The trajectories for each decade were sorted into eight clusters organized clockwise from North. The revised clustering program used the same approach as the program distributed with HYSPLIT, finding the minimum spatial variance between clusters.