29 h May 2025 Vicky Lucas
Ho Topics:
Machine Lea ning o Wea he Fo ecas ing &
Hea wa e P edic ion o Medium Range & Beyond
A Guide o Me eo ologis s – Bu n A e Reading
Why ead his epo ?
W i en o ope a ional o esea ch me eo ologis s e e ywhe e, his is a guide on how machine
lea ning is being used in wea he p edic ion, wi h a ocus on he challenge o ex eme hea . Cu ing
h ough he hype – i ’s an app oachable e iew on wha is happening and wha o wa ch. I you a e
wonde ing whe e hings a e heading, keep eading – some o he ie ies issues a e s ill un esol ed.
1. Deg ees o change: in oduc ion ........................................................................................... 2
2. Me cu y ising: ch onology o machine lea ning wea he models ....................................... 3
3. On he ho izon: ML in hea wa e o ecas ing & beyond medium ange ................................ 7
4. The hea is on: la es de elopmen s ...................................................................................... 9
5. Flashpoin : endu ing challenges .......................................................................................... 10
6. Glossa y: ligh elie ............................................................................................................. 11
Ho Topics: Machine Lea ning o Wea he Fo ecas ing & Hea wa e P edic ion Medium Range & Beyond
29 h May 2025 UBC Clima e Solu ions Resea ch Collec i e – Solu ions Schola s P og am 2024/25 Vicky Lucas
1. Deg ees o change: in oduc ion
Machine lea ning (ML) is eshaping me eo ological modelling, o e ing new ways o emula e,
imp o e o eplace a ange o wea he and clima e modelling (Chan y e al., 2021). In jus a ew
yea s echniques ha e mo ed om ocused applica ions, such as ada nowcas ing, o ull global
models using machine lea ning o wea he p edic ion (MLWP) ha i al, and in some cases
ou pe o m, adi ional nume ical wea he p edic ion (NWP) sys ems. The shi is no jus in who
builds hese models, o en echnology companies, bu in how o ecas s a e p oduced. MLWP does
no sol e physical equa ions, bu lea ns pa e ns in da a, based on echniques used in o he machine
lea ning applica ions such as ecommenda ion engines, image gene a ion and La ge Language
Models (LLMs).
Many MLWP sys ems a e ained on his o ical and cons uc ed wea he da ase s, e y o en
eanalysis – and especially ERA5 (Clima e Da a S o e). MLWP model aining equi es signi ican
da a, ime and ene gy, bu once ained, ope a ional ML models can p oduce o ecas s in seconds o
minu es on minimal ha dwa e compa ed wi h NWP supe compu e s. This speed opens up
possibili ies o la ge ensembles, apid upda es, and hyb id wo k lows.
Some hing o bea in mind is ha his is a domain in lux, no a chi ec u e clea ly p e ails. G oups a e
expe imen ing wi h con igu a ions, p og essi ely leap ogging hei own and o he s’ achie emen s.
While many MLWP models appea in pee - e iewed jou nals, some emain as p e-p in s – including
e e ences in his epo (e.g., a Xi ), and some make open-sou ce e sions a ailable.
As de elopmen s con inue, he ple ho a o global gene al MLWP models can be di ided by hei
aining egimes (Shi e al., 2025 – a Xi ):
• De e minis ic p edic i e lea ning, as used in Fou Cas Ne , uses supe ised aining – labelled
da a whe e each inpu comes wi h a co esponding ou pu – a lea n ela ionship. These
models minimise o ecas e o , a e as o un, bu accumula e e o s o e longe lead imes.
• P obabilis ic gene a i e modelling, such as GenCas , uses di usion o gene a e ensemble
o ecas s, o e ing unce ain y modelling wi hou explici pe u ba ions o NWP ensembles.
Gene a i e models lea n he dis ibu ion o he aining da a, hen gene a es new samples.
• Founda ion models, such as Au o a, p e- ained on ex emely la ge and di e se da ase s o
ul ill a b oad ange o asks, ans e ing lea ning o o he da ase s ei he wi h ine- uning
on new da a o p omp ing e.g., om wea he o ai quali y modelling.
This b oad aining delinea ion is a amewo k o unde s and ML in me eo ology. In some ins ances,
deep lea ning, a subse o ML, would be a mo e accu a e e m, bu ML is used he e o simplici y. A
glossa y o ML e ms is included a he end o his epo – when new e ms a e used a hype link o
he glossa y is p o ided.
This epo ollows how MLWP is shi ing om esea ch o eal-wo ld use. I s a s by acing he
ecen p oli e a ion o global ML models, in oducing some o he no able pa icipan s and a li le on
how hey wo k. Nex , is an explo a ion on wha is happening in o ecas ing hea wa es and ex ending
he o ecas ing window beyond he medium ange and in o sub-seasonal o seasonal (S2S). The inal
sec ions look ahead o whe e he ield is going – concluding wi h un esol ed ensions and challenges.
Ho Topics: Machine Lea ning o Wea he Fo ecas ing & Hea wa e P edic ion Medium Range & Beyond
29 h May 2025 UBC Clima e Solu ions Resea ch Collec i e – Solu ions Schola s P og am 2024/25 Vicky Lucas
2. Me cu y ising: ch onology o machine lea ning wea he models
The apid de elopmen o gene al global machine lea ning wea he models e lec s echnology
company e o s in me eo ological modelling and capi alising on echniques, such as used in La ge
Language Models e.g., Cha GPT ( ans o me s) o image gene a ion (di usion). Since 2022 MLWP
models ha e made headlines on ou pe o ming espec ed ope a ional NWP. These ML models we e
op imised o minimising a e age e o s, bu la e models ha e swi ched o gene a i e modelling and
wo ypes o his app oach, GenCas and Au o a, a e discussed in mo e de ail la e .
A imeline o gene al global ML models 2022 o 2025 is shown in Figu e 1, and a glossa y o ML e ms
used is p o ided a he end o his epo .
Fou Cas Ne (Pa hak e al., 2022 – a Xi ) was he
i s gene al-pu pose de e minis ic MLWP model
based on a neu al ne wo k ha i aled he accu acy
o ope a ional NWP. Fou Cas Ne was ained on
10TB o da a, emula ing a mosphe ic dynamics
lea n om ERA5 eanalysis. I can be un globally in
seconds a 0.25° esolu ion. The b eak h ough was
wo old. Fi s ly, using he Fou ie Neu al Ope a o
app oach (Li e al., 2021 – a Xi ) which lea ns
mapping be ween en i e spa ial ields ( unc ion
space) a he han indi idual and neighbou ing g id
poin s. I s Fou ie a chi ec u e cap u es bo h low-
and high- equency componen s, e icien ly
ep esen ing global and local s uc u es. Second,
he in eg a ion o a Vision T ans o me enables he
model o lea n long- ange spa ial dependencies.
The Vision T ans o me was a landma k pape on
sel -a en ion, which allows ne wo ks o p io i ize
ele an in o ma ion when lea ning; published by
Google in 2017 and used in LLMs such as Cha GPT –
i is one o he mos ci ed pape s his cen u y
(Pea son e al., 2025). Toge he , hese inno a ions
enabled as lea ning, high esolu ion, and e icien
compu a ion.
G aphCas (Lam e al., 2023) appea ed in p e-p in s soon a e , la e published in Science, applying
G aph Neu al Ne wo ks o global o ecas ing. GNNs ep esen he a mosphe e as an icosahed al
mesh, enabling uni o m spa ial esolu ion. Nodes (g id poin s) a e connec ed by edges which de ine
he low o in o ma ion be ween all nodes – allowing he model o lea n eleconnec ions ega dless
o dis ance o inpu g id egula i y. Fo ecas ing o e 1,000 a iables up o 10 days ahead, G aphCas
achie ed g ea e accu acy han ECMWF IFS con en ional NWP model – including a 10 day lead imes
o he op 2% o ho days o e land (Lam e al., 2023).
P e ious models had aken hei names om hei a chi ec u e, bu Pangu-Wea he changed ha –
named a e a p imo dial my hical being who sepa a ed hea en and ea h and became geog aphic
ea u es. This de e minis ic model in oduced wo new app oaches (Bi e al., 2023). Fi s ly, i used a
h ee-dimensional app oach, using e ical s uc u e ia p essu e le el da a. Secondly i ained
sepa a e models o 1, 3, 6 and 24-hou lead imes, which educed e o s o e long lead imes.
Figu e 1: imeline o selec ed ML wea he models
Ho Topics: Machine Lea ning o Wea he Fo ecas ing & Hea wa e P edic ion Medium Range & Beyond
29 h May 2025 UBC Clima e Solu ions Resea ch Collec i e – Solu ions Schola s P og am 2024/25 Vicky Lucas
Th oughou hese ad ances ECMWF, he Eu opean Cen e o Medium-Range Wea he Fo ecas s –
bo h a esea ch ins i u e and ope a ional se ice – has been cen al in he applica ion o ML. The
ECMWF ope a ional NWP, he IFS HRES, is o en used as he benchma k model compa ison, o he
ensemble IFS ENS. ECMWF de eloped i s own de e minis ic g aph-based (GNN) sys em, AIFS (Lang
e al., 2024 – a Xi ), and an ensemble AIFS-ENS based on di usion (Figu e 1) (Lang e al., 2024).
Di usion is he same app oach used o image gene a ion and in a la e MLWP model, GenCas –
ea u es in he nex sec ion o his epo , whe e di usion will be discussed in mo e de ail. The AIFS-
ENS was made ope a ional in ea ly 2025, alongside he longs anding con en ional NWP he IFS.
ECMWF displays hi d-pa y MLWP ou pu s in eal- ime and e alua es model pe o mance. Figu e 2
shows compa a i e pe o mance h ough he oo mean squa ed e o o wo-me e ai empe a u e
o he no he n ex a- opics, o win e 2024/25 – showing ha Au o a consis en ly achie es he
g ea es accu acy on ha b oad me ic – ou pe o ming he ope a ional NWP, IFS. A ange o cha s
and pe o mance sco es a e a ailable om he ECMWF cha ca alogue.
Figu e 2: RMSE 2m empe a u e, win e 24/25 no he n ex a- opics - consis en ly bes pe o mance om Au o a
IFS (ECWMF ope a ional NWP) ( ed) and i e MLWP models: AIFS (ECMWF) (b own), Fou Cas Ne (g een), G aphCas (cyan),
Pangu-Wea he (da k blue) & Au o a (o ange) – a ailable om ECMWF
Ho Topics: Machine Lea ning o Wea he Fo ecas ing & Hea wa e P edic ion Medium Range & Beyond
29 h May 2025 UBC Clima e Solu ions Resea ch Collec i e – Solu ions Schola s P og am 2024/25 Vicky Lucas
ClimaX (Nguyen e al., 2023 – a Xi ) was he i s ounda ion model, ained on uly as and
he e ogeneous da ase s including ERA5 and he clima e model ou pu o CMIP6. The app oach
in oduced a sha ed ans o me encode p e- ained, and which can be ine- uned o mo e speci ic
asks. ClimaX is capable o global and egional o ecas s, also sub-seasonal and clima e p ojec ions,
and downscaling. Simila ly, P i h i WxC (Schmude e al., 2024 – a Xi ) is a la e ounda ion model, a
collabo a ion be ween IBM and NASA.
Ano he i s , Aa d a k, has ecen ly been published in Na u e (Allen e al., 2025). The sys em has
wo no able ea u es, i s ly i can un on a desk op compu e in minu es (Tu ne , 2025). Secondly,
whils i has been ained on ERA5, i does no need any NWP p oduc s o un, i only equi es
obse a ions as inpu such as sa elli e and wea he s a ion da a. This end- o-end sys em, ully
obse a ion-d i en, has been an aspi a ion in he ield (McNally e al., 2024). Aa d a k shows s ong
pe o mance agains ope a ional NWP, e en wi h a coa se 1.5 deg ee ou pu and using jus 10% o
inpu s o exis ing NWP sys ems (Tu ne , 2025). These ea u es make Aa d a k aluable o
applica ions like disas e p epa edness in low- esou ce se ings.
The de ails: GenCas & Au o a
To illus a e ecen a chi ec u es, wo models a e de ailed he e, GenCas and Au o a (Figu e 1).
GenCas is a p obabilis ic gene a i e model o medium ange 15 day o ecas ing (P ice e al., 2025)
while Au o a is a ounda ion model which can be uned o mul iple applica ions including o ecas ing
ai quali y and opical cyclone acks (Bodna e al., 2024 – a Xi ).
GenCas is a gene a i e model – in aining i lea ns he unde lying da a dis ibu ion and hen
p oduces new ealis ic samples. This gene a i e unc ion comes om using a di usion p ocess,
based on noise. In aining, o p edic condi ions 12 hou s ahead (T+12), GenCas akes wo inpu
ields om ERA5, e ec i ely T-12 and T+0. Ra he han o ecas ing T+12 di ec ly, i compu es he
esidual (i.e., he change be ween T–12 and T+0) and adds scaled noise o c ea e he i s guess.
GenCas hen i e a i ely emo es noise o e mul iple s eps. By pe u bing he endency a he han
he absolu e s a e, i p ese es cohe en a mosphe ic s uc u es. GenCas was p e- ained a 1°
esolu ion on 40 yea s o ERA5 da a, hen ine- uned a 0.25°.
In o ecas mode, known as in e ence in ML, GenCas p edic s each T+12 using he wo mos ecen
s a es, con inuing un il i eaches T+360 (15 days). This ype o olling p edic ion is called
au o eg ession in ML. Gene a ing an ensemble is s aigh o wa d due o he di usion app oach;
each ensemble membe s a s om a di e en scaled noise sample. 50 ensemble membe s we e
chosen o ma ch ECMWF IFS ENS – a lo mo e could be p oduced. While di usion models a e
compu a ionally in ensi e—each s ep in ol es mul iple denoising passes—GenCas can p oduce a
ull 15-day o ecas in minu es pe membe . I s G aph Neu al Ne wo k (GNN), adap ed om
G aphCas (discussed ea lie ), cap u es and suppo s spa ial dependencies ac oss ime s eps.
GenCas ou pe o med IFS ENS, including ex eme empe a u es, he 99.9 h and 99.99 h pe cen iles,
ou o 15 days (P ice e al., 2025).
In con as , Au o a is a ounda ion model – he e m ounda ion model came om S an o d in 2021 –
whe e he model is a common basis o many ask speci ic adap a ions. Au o a is p e- ained on
mo e han one million hou s o global Ea h sys em da a, including wea he eanalyses, ope a ional
o ecas s, and he CMIP6 clima e simula ions. I con ains 1.3 billion pa ame e s and uses a 3D Swin
T ans o me , an a chi ec u e adap ed om compu e ision. The Swin, shi ed window, ans o me
imp o es e iciency by di iding da a in o local windows and hen shi ing hem be ween laye s o
cap u e b oade spa ial and e ical dependencies. Au o a also employs pe cei e -based encode s
Ho Topics: Machine Lea ning o Wea he Fo ecas ing & Hea wa e P edic ion Medium Range & Beyond
29 h May 2025 UBC Clima e Solu ions Resea ch Collec i e – Solu ions Schola s P og am 2024/25 Vicky Lucas
and decode s which allow he model o inges a ying ypes and shapes o da a, ans o ming hem
o a common in e nal o ma o combine di e en inpu da a.
These a chi ec u al choices allow Au o a o model complex Ea h sys em dynamics ac oss domains.
A e p e aining, i can be ine- uned o asks like 10-day global wea he o ecas ing, 5-day ai
quali y p edic ion, ocean wa e modelling, and opical cyclone acking. Au o a (Bodna e al., 2024
– a Xi ) p oduces 10-day global o ecas s in unde a minu e and 5-day ai quali y p edic ions in
seconds.
Summa y
F om he as un imes o Fou Cas Ne o he lexible uning o Au o a and he ensemble unce ain y
quan i ica ion o GenCas , MLWP is app oaching, and by some me ics exceeding NWP pe o mance,
all wi h signi ican ly educed ope a ional compu a ional cos . The enable s o ha p og ess include:
• Open access o high- esolu ion da ase s such as ERA5,
• C oss-disciplina y a chi ec u e design (e.g. ans o me s, di usion),
• Collabo a ion be ween esea ch ins i u es, ope a ional cen es, and ech companies,
• Access o MLWP on Gi Hub and Hugging Face – o sha ing da a, code and documen a ion.
Fo u he eading, Chen e al. (2023) o e s a good o e iew o me hods and a chi ec u es, while
Zhang e al. (2025) p o ides an up- o-da e e iew o knowledge, me hods, ends and challenges.
Waqas e al. (2024) o e s a sys ema ic e iew o AI in eg a ed wi h NWP. Fo mo e on MLWP
compa a i e pe o mance, Google Resea ch has a sco eca d o se e al models a Wea he Bench2.
Ho Topics: Machine Lea ning o Wea he Fo ecas ing & Hea wa e P edic ion Medium Range & Beyond
29 h May 2025 UBC Clima e Solu ions Resea ch Collec i e – Solu ions Schola s P og am 2024/25 Vicky Lucas
3. On he ho izon: ML in hea wa e o ecas ing & beyond medium ange
Gene al MLWP a medium ange & S2S
MLWP is ex ending skill in o he medium and sub-seasonal ange (15–46 days). Two de e minis ic
gene al pu pose ans o me models, FuXi and FengWu ha e mo ed o a chi ec u es wi h ensemble
gene a ion o o ecas s ou o six weeks in S2S e sions. FuXi-S2S demons a es imp o ed skill o e
ECMWF IFS S2S o se e al me ics, including 2m empe a u e and Madden-Julian Oscilla ion (MJO)
p edic ion up o day 36 (Chen e al., 2024). FengWu-W2S uses ocean–a mosphe e–land coupling
cons ain s o physical consis ency (Ling e al., 2024). The esul s o 2m empe a u e show
compa able bu g ea e skill han ECMWF S2S and FuXi-S2S a h ee o six weeks. FengWu-W2S can
p edic he MJO o 37 days, sligh ly longe han ECMWF S2S o FuXi-S2S, and p edic s he No h
A lan ic Oscilla ion wi h g ea e skill han ECMWF S2S a ou o six weeks.
Clima e
ML models a e also being es ed in clima e applica ions. Neu alGCM (Kochko e al., 2024) is a
hyb id model ained on ERA5 ha pe o ms well o 10–15 day o ecas s and can simula e clima e
me ics o e mul iple decades when o ced wi h p esc ibed sea su ace empe a u es. The model is
no coupled o land o ocean and does no include g eenhouse gases o ae osols. Neu alGCM does
no show a clea end o inc easing e o when ini ialised u he in o he u u e om he aining
da a. ‘Ou esul s p o ide s ong e idence o he dispu ed hypo hesis ha lea ning o p edic sho -
e m wea he is an e ec i e way o une pa ame e isa ions o clima e.’ (Kochko e al., 2024).
ML imp o ing local & egional NWP o ex eme hea
Machine lea ning is being used o imp o e hea wa e o ecas ing a local scales and o ex end skill in
o ecas ing su ace ex emes. A he u ban le el, ML has been used o bias co ec ion and
downscaling. A ecen London s udy (Blunn e al., 2024) used ML o combine high esolu ion NWP
ou pu wi h ci izen science wea he s a ion obse a ions du ing hea wa es. ML educed mean
absolu e empe a u e e o by 11%, iden i ying la en hea lux as he mos impo an p edic o o
empe a u e bias. In egional modelling, ML can suppo sensi i i y analysis o NWP physical
schemes. One example om Aus alia (Reddy e al., 2023) used ML o iden i y, o wen y ou
pa ame e s, he wo key d i e s o modelled su ace empe a u e and ela i e humidi y p edic ions:
he sho wa e adia ion sca e ing pa ame e and sa u a ed soil wa e con en mul iplie . These wo
s udies show complemen a y uses o ML in imp o ing con en ional NWP.
MLWP ex ending lead imes o ex eme hea p edic ion
Pape s om 2022 o 2025 explo ing he use o ML app oaches o ex ending he lead imes o
hea wa e p edic ions is de ailed in Table 1, and se e al discussed below. The able also includes
ecen pape s on explainable AI (XAI) discussed in he nex sec ion, and wo e iew pape s.
Among he mo e ecen and longe - ange examples is he Wei ich-Bene e al. (2023) s udy, which
applied bo h linea eg ession and andom o es s o o ecas summe hea wa es in Cen al Eu ope
a lead imes o 1–6 weeks. Key p edic o s included 500 hPa geopo en ial heigh , soil mois u e, and
sea su ace empe a u es. While model pe o mance declined wi h inc easing lead ime, he
machine lea ning ou pu s ou pe o med pe sis ence and clima ology. Beyond wo weeks, hei
o ecas s we e as skil ul as he ECMWF sub seasonal ensemble-mean hindcas o he egion.
Lopez-Gomez e al. (2023) explo ed ML con igu a ions ained speci ically on ex eme empe a u es
— a he han all empe a u es. This adjus men led o imp o ed pe o mance o e pe sis ence and
compa able skill wi h ECMWF S2S a e day 14. The MLWP could o ecas ou -o -sample e en s, bu
s uggled wi h ex emely ho days, abo e he 95 h pe cen ile.
Ho Topics: Machine Lea ning o Wea he Fo ecas ing & Hea wa e P edic ion Medium Range & Beyond
29 h May 2025 UBC Clima e Solu ions Resea ch Collec i e – Solu ions Schola s P og am 2024/25 Vicky Lucas
In a case s udy o he No h Ame ican Paci ic No hwes 2021 hea wa e, he anomalous
empe a u es we e only p edic ed 2 o 5 days p io , al hough a abou a mon h ahead he MLWP was
simila in s uc u e o he ECMWF ensemble. Duan e al. (2025), apply he Neu alGCM, discussed
abo e, o explo e he Paci ic No hwes hea wa e, which i eplica es.
Xie e al. (2024), used a Con olu ional Neu al Ne wo k – an app oach ha inds local spa ial
co ela ions in g idded da a – o o ecas hea wa es in China up o 30 days ahead. The me hod
il e ed ou high- equency signals (<10 days) and isola ed he 10–90 day low- equency backg ound
s a e, using inpu s om NCEP/NCAR eanalysis and pola -o bi ing sa elli e da a. While o ecas skill
was lowe up o 20 days, he model ou pe o med bo h he China Me eo ological Adminis a ion
NWP and ECMWF S2S ensemble mean be ween days 20 and 30 when a e aged ac oss he coun y.
Table 1: summa y o pape s on ML app oaches applied o hea ex emes om 2022 o 2025
Re e ence & link
Da a o model
A ea
Technique
Lead ime o
pu pose
Ennis e al., 2025
G aphCas , FuXi,
Pangu, GEFS
USA
MLWP – NWP compa ison
Up o 20 days
Lo o e al., 2025
CESM (NCAR)
F ance
Mul iple ML models, also
XAI s udy
XAI
Sha iq e al.,
2025
5 yea s wea he
obse a ions
Laho e
XAI s udy, LSTM bes
pe o mance
1-3 days, XAI
Camps-Valls e
al., 2025
Mul iple
Global
Re iew o me hods
Re iew pape
Duan e al., 2025
ERA5
Paci ic
No hwes
Neu alGCM and ensemble
compa ison E3SM NWP
6 days
Xie e al., 2024
NCEP/NCAR
eanalysis
China
CNN wi h il e ing
Up o 30 days
Wei ich Bene e
al., 2023
ECMWF
Eu ope
Linea and Random Fo es
1-6 weeks
Lopez Gomez e
al., 2023
ERA5
Global
Mul iple ML models
Up o 28 days
Salcedo-Sanz e
al., 2023
Mul iple
Global
Re iew o me hods
Re iew pape
MLWP o ex emes – open ques ions
Despi e encou aging case s udies, ecen e iews cau ion ha MLWP is no eady o ope a ional
o ecas ing o ex emes. Oli e i & Messo i (2024) highligh h ee key limi a ions: (i) mos models a e
uned o a e age o ecas skill, no ex emes; (ii) a chi ec u es a e no op imised o he limi ed da a
a ailable on a e e en s; and (iii) assump ions abou e o dis ibu ions a e o en oo simplis ic. They
also no e ha leading global MLWP lack alida ion o ex eme e en p edic ion.
Salcedo-Sanz e al. (2023) p o ide a his o y and b oade c i ique, ocused on ex eme e en s and a
li e a u e e iew o hea wa es. They also emphasise he limi ed da a a ailable o aining on a e
e en s and a gue ha mos cu en ML sys ems lack in eg a ion wi h physical clima e knowledge o
achie e eliable p edic ions. The au ho s call o he use o mul iple eanalysis da ase s, imp o ed
model anspa ency h ough explainable AI (XAI), and g ea e ocus on compound and concu en
ex emes. Simila ly, he e iew pape o Camps-Valls e al., (2025) discusses he hu dle o limi ed
aining da a, and o deploying unde s andable models – needed o gaining us . Many au ho s
unde sco e ha MLWP equi es consis en and anspa en de elopmen o ex emes.
Ho Topics: Machine Lea ning o Wea he Fo ecas ing & Hea wa e P edic ion Medium Range & Beyond
29 h May 2025 UBC Clima e Solu ions Resea ch Collec i e – Solu ions Schola s P og am 2024/25 Vicky Lucas
4. The hea is on: la es de elopmen s
As MLWP ma u es, esea che s a e esponding o c i iques — including he alida ion o ex eme
e en o ecas s. Recen wo k e alua ed h ee global MLWP models (Fou Cas Ne , Pangu-Wea he ,
G aphCas ) alongside ECMWF IFS HRES o e en s including he 2021 Paci ic No hwes hea wa e
(Pasche e al., 2025). While ML models cap u ed he b oad s uc u e o he e en , hey
unde pe o med when spa ially and empo ally agg ega ed. Fou Cas Ne had he la ges e o s;
Pangu-Wea he ended o o e es ima e a ec ed a eas o a gi en h eshold. A lead imes unde a
week, he IFS HRES emained he mos accu a e. No ably, MLWP ha e ew a iables – lacking su ace
ela i e humidi y – and he au ho s indica ed he need o sea su ace empe a u e and soil mois u e
inpu s o longe lead imes. A sepa a e s udy examined 60 U.S. hea wa es and ound G aphCas
consis en ly ou pe o med Pangu-Wea he and he US GEFS ensemble ac oss mos egions wi h
ERA5 used as g ound u h (Ennis e al., 2025 – a Xi ). The open-sou ce na u e o some MLWP
models allows o his independen analysis and e i ica ion – which is likely o con inue – assis ed by
he Wea he Bench2 (Rasp e al., 2023 – a Xi ) da abase o ex eme e en cases, (also Clima eBench
(Wa son-Pa is e al., 2022)) and ECMWF adding MLWP o ecas e alua ion o i s da abase o
ex eme cases (Magnusson, 2023).
Explainabili y emains a majo conce n. While s udies by Lo o e al. (2025), Sha iq e al. (2025), and
Wei e al. (2025 – a Xi ) ha e applied explainable AI (XAI) ools o hea o ecas ing, hese o en ely
on simpli ied models o componen s. As Lo o no es, Con olu ional Neu al Ne wo ks – he ea lie
p edic i e app oach o MLWP sui ed o g idded da a and used e.g., by G aphCas – ‘a e black boxes
e en when using XAI ools’. Mo e anspa en me hods, hyb id models combining ML wi h physical
cons ain s, a e impo an o scien i ic us and likely o be so o public us .
A po en ial end is he in eg a ion o la ge language models (LLMs) o suppo o ecas
in e p e a ion and decision-making. Wild i eGPT, de eloped as a p oo -o -concep , combined
ex eme wea he p ojec ions, li e a u e, and use p omp s o gene a e geospa ially con ex ualized
esponse guidance (Xie e al., 2024 – a Xi ). Designed o be use -cen ic and isually in eg a ed, i
ou pe o med gene ic LLMs in en case s udies. A simila ‘Hea wa eGPT’ could help b idge he gap
be ween echnical NWP, ML ou pu s, ope a ional needs and ac ions, and public- acing messaging.
Looking ahead, he ield is s ill e ol ing and di e si ying. Aa d a k, capable o gene a ing o ecas s
di ec ly om obse a ions, ep esen s a shi in a chi ec u e – able o include some measu emen s
oo complex o NWP assimila ion (McNally e al., 2024). And new sys ems egula ly appea such as
he open sou ce Wea he Mesh-3 (Du e al., 2025) by an a mosphe ic sensing company, WindBo ne.
The S2S imescales and beyond will be a domain o in e es ; jus as MLWP can ou pe o m NWP a
sho o medium ange (1 o 15 days) seasonal p edic ions will imp o e o MLWP – including global
seasonal o decadal p edic ions such as ACE2 (Wa -Meye e al., 2024, Ken e al., 2025 – a Xi ).
Ano he ecen de elopmen is om NVIDIA – Ea h-2 is hei comme cial digi al win which can be
used o kilome e scale wea he and clima e p edic ions, i is ocused on isk and is a business
o e ing (NVIDIA, 2025). Es ablishing business models could be a nex s ep o o he o ganisa ions
ha ha e in es ed esou ces o e he las ew yea s.
As alida ion con inues, explainabili y e ol es and new models appea , he impo ance o ML in
me eo ological p edic ion looks se o emain. P og ess could be inc emen al o adical depending
on pe o mance and esou cing p essu es – bu going oo a oo as could unde mine us and
accep ance – om science and om he public.
