BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Online-Learned Neural Network Chemical Solver for Stable\, Fast\, 
 and Long-Term Global Simulations of Atmospheric Chemistry - Dr Makoto Kelp
 \, NOAA Climate and Global Change Postdoctoral Fellow\, Dept of Earth Syst
 em Science\, Stanford University
DTSTART:20240206T160000Z
DTEND:20240206T170000Z
UID:TALK209704@talks.cam.ac.uk
CONTACT:Annabelle Scott
DESCRIPTION:Global models of atmospheric chemistry are computationally exp
 ensive. A bottleneck is the chemical solver that integrates the large-dime
 nsional coupled systems of kinetic equations describing the chemical mecha
 nism. Machine learning (ML) could be transformative for reducing the cost 
 of an atmospheric chemistry simulation by replacing the chemical solver wi
 th a faster emulator. However\, past work found that ML chemical solvers e
 xperience rapid error growth and become unstable over time. In this talk\,
  I will present the culmination of several years of research focused on de
 veloping ML methods for atmospheric chemistry simulations. We started by e
 stablishing stable emulation in 0-D box models and then progressed to achi
 eving for the first time a stable full-year global chemical transport mode
 l (CTM) simulation of atmospheric chemistry using ML solvers. The ML solve
 r gains five-fold speedup in computational performance over the reference 
 Fortran solver during a CTM simulation. We show that online training of th
 e ML solver synchronously with the CTM simulation produces considerably mo
 re stable results than offline training from a static data set of simulati
 on results. Although our work represents an important step for using ML so
 lvers in global atmospheric chemistry models\, more work is needed to exte
 nd it to large chemical mechanisms and to reduce errors during long-term c
 hemical aging.
LOCATION:Drum Building\, Madingley Rise Site\, West Cambridge and on zoom:
   https://zoom.us/j/6708259482?pwd=Qk03U3hxZWNJZUZpT2pVZnFtU2RRUT09
END:VEVENT
END:VCALENDAR