BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:Europe/Stockholm
X-LIC-LOCATION:Europe/Stockholm
BEGIN:DAYLIGHT
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
TZNAME:CEST
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
TZNAME:CET
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20210916T132455Z
LOCATION:
DTSTART;TZID=Europe/Stockholm:20210706T173000
DTEND;TZID=Europe/Stockholm:20210706T190000
UID:submissions.pasc-conference.org_PASC21_sess182_post105@linklings.com
SUMMARY:P01 - Neural Nets as an Aid for Constructing Tangent-Linear and Ad
 joint Models
DESCRIPTION:Poster\n\nP01 - Neural Nets as an Aid for Constructing Tangent
 -Linear and Adjoint Models\n\nHatfield, Chantry, Dueben\n\nLinearised nume
 rical models have a number of uses in computational science, notably in nu
 merical weather prediction. Tangent-linear models allow us to evolve pertu
 rbations forwards in time, whereas adjoint models allow us to propagate gr
 adients backwards in time. Both are essential for the incremental 4D-Var d
 ata assimilation algorithm used operationally at many weather prediction c
 entres, including the European Centre for Medium-Range Weather Forecasts, 
 to generate the initial conditions for forecasts. However, linear models a
 re notoriously difficult to develop and maintain, in terms of code complex
 ity, and are often considerably more computationally expensive than the or
 iginal nonlinear model. In this poster I will outline how machine learning
  could aid in the construction and evaluation of linearised numerical mode
 ls. Several researchers have recently proposed using neural networks to re
 place components of weather prediction models, in particular the parametri
 zation schemes for unresolved physics. If a parametrization scheme in the 
 nonlinear model could be successfully emulated then a linearised version o
 f the scheme could be generated automatically, by taking advantage of the 
 inherent differentiability of the neural network. I will describe here the
  basic mathematics underlying this idea and present some simple numerical 
 results demonstrating its viability.
END:VEVENT
END:VCALENDAR