AI, a new asset for weather forecasters

Weath­er fore­cast­ing is now a com­mon prac­tice for which met­eor­o­lo­gists require atmo­spher­ic sim­u­la­tion mod­els. Météo-France reg­u­larly rein­forces its intens­ive com­put­ing capa­cit­ies, qual­ity of numer­ic­al weath­er pre­dic­tion mod­els, and the use of obser­va­tion data to cal­ib­rate them.

The research car­ried out at CNRM (Centre nation­al de recherches météoro­lo­giques), in col­lab­or­a­tion with sev­er­al nation­al and inter­na­tion­al part­ners, focuses on under­stand­ing the atmo­spher­ic envir­on­ment and its inter­ac­tions with con­tin­ent­al and ocean­ic sur­faces. The CNRM devel­ops invest­ig­at­ive tools for observing and improv­ing our under­stand­ing of pro­cesses at obser­va­tion sites. It also works to devel­op mod­els for numer­ic­al sim­u­la­tions our ever-chan­ging atmo­sphere and its inter­faces, doing so on a scale of hours, days or much longer. These mod­els oper­ate on mul­tiple spa­tial scales, from sev­er­al tens of metres to hun­dreds of kilo­metres – or even on a glob­al scale.

These weath­er mod­els are used to fore­cast the weath­er and study cli­mate change through cli­mate pro­jec­tions. There are sev­er­al mod­el­ling tools, such as Arpège and Arome. Arpège is used to sim­u­late the atmo­sphere of the entire plan­et, cov­er­ing Europe with a mesh size of around 5 km and the rest of the globe with mesh sizes ran­ging from 5–24 km. Arome cov­ers a domain lim­ited to met­ro­pol­it­an France and neigh­bour­ing coun­tries with a hori­zont­al res­ol­u­tion of 1.3 km, and is also deployed over over­seas regions.

Arome has been designed to oper­ate over a part of the world only, which may be centred on some over­seas ter­rit­or­ies and oth­er loc­a­tions of interest. The mod­el is used to improve the short-term fore­cast­ing of haz­ard­ous phe­nom­ena such as heavy Medi­ter­ranean rain­fall (‘Medi­ter­ranean epis­odes’), severe thun­der­storms, fog or urb­an heat islands dur­ing heat waves. It allows sim­u­la­tions to be car­ried out with very fine res­ol­u­tion, by sli­cing the atmo­sphere into small cubes (known as ‘meshes’) to solve the equa­tions gov­ern­ing the phys­ic­al pro­cesses in the atmo­sphere and at the interfaces. 

The raw res­ults of these sim­u­la­tions must be pro­cessed to pro­duce fore­casts in an oper­a­tion­al frame­work. In this con­text, CNRM is increas­ingly using advanced stat­ist­ic­al meth­ods, includ­ing arti­fi­cial intel­li­gence (AI), to com­bine past fore­casts with what has been observed, to some­how adjust the mod­els so that they are as rel­ev­ant as pos­sible to the observations.

AI is a meth­od­o­lo­gic­al tool that has been emer­ging for sev­er­al years. It can speed up the exe­cu­tion of a weath­er mod­el and thus reduce its com­pu­ta­tion­al cost. For example, one of the most expens­ive aspects of mod­el exe­cu­tion is solv­ing the com­plex equa­tions that gov­ern radi­at­ive trans­fer through the atmo­sphere, through clouds, and the inter­ac­tions of light rays with the land or ocean sur­face. Future research could speed up mod­el exe­cu­tion by sub­sti­tut­ing some expli­cit res­ol­u­tions of the phys­ic­al equa­tions with res­ults obtained through machine learn­ing. Some work goes so far as to sug­gest the out­right replace­ment of fore­cast­ing mod­els solv­ing phys­ic­al equa­tions with deep learn­ing emu­la­tion of fore­cast­ing systems.

Anoth­er applic­a­tion of AI is the post-pro­cessing of mod­el­ling res­ults for weath­er fore­casts. This pro­cess allows the cor­rec­tion of cer­tain sys­tem­at­ic errors in the out­put of the mod­els to make the fore­cast more con­sist­ent with obser­va­tions in sim­il­ar cir­cum­stances. It also makes it pos­sible to detect struc­tures cor­res­pond­ing to phe­nom­ena at stake, which is all the more dif­fi­cult as fore­casts are increas­ingly made with­in the frame­work of ensemble fore­casts, mak­ing it pos­sible to bet­ter account for the uncer­tain­ties of the fore­casts. Among the work in pro­gress at the CNRM, we should men­tion, for example, work aimed at improv­ing the detec­tion of par­tic­u­larly viol­ent thun­der­storms in the res­ults of numer­ic­al sim­u­la­tions, which are extremely import­ant phe­nom­ena to detect but dif­fi­cult to fore­cast¹.

One approach to pro­du­cing increas­ingly rich weath­er fore­casts is to reduce the spa­tial scale on which they oper­ate, thereby rep­res­ent­ing phys­ic­al pro­cesses that can be solved on a smal­ler scale. As men­tioned earli­er, the Arome fore­cast sys­tem, for example, cur­rently oper­ates at a hori­zont­al res­ol­u­tion of 1.3 km. A high­er res­ol­u­tion would allow even smal­ler scale pro­cesses to be rep­res­en­ted, for example those lead­ing to very loc­al­ised high-impact thun­der­storms. How­ever, a high-res­ol­u­tion mod­el does not guar­an­tee that the loc­a­tion of the most intense thun­der­storms can be accur­ately iden­ti­fied. Thun­der­storms are scattered phe­nom­ena, not wide­spread events over large parts of the ter­rit­ory, as are heat waves, and are there­fore more dif­fi­cult to pre­dict in terms of their occur­rence, loc­a­tion, and intens­ity. Obser­va­tions play a fun­da­ment­al role in ini­tial­iz­ing weath­er fore­cast­ing sys­tems, with meth­ods known as ‘data assimilation’.

Work is under­way to exploit more innov­at­ive obser­va­tions such as so-called ‘oppor­tun­ity’ data, col­lec­ted by par­ti­cip­at­ory tools such as indi­vidu­al con­nec­ted weath­er sta­tions. These data are col­lec­ted by the com­pany that mar­kets these instru­ments and allow us to enrich our know­ledge even if the intrins­ic qual­ity is not at the same level as that of the con­ven­tion­al network.

There are also devel­op­ments in the obser­va­tions obtained from air­craft and ships, as well as from bal­loons, equipped with probes, which are sent into the atmo­sphere sev­er­al times a day. And then, of course, we have many obser­va­tions made by satel­lites, either geo­sta­tion­ary or defil­ing, which provide a view on very large spa­tial scales. The former provides us with a per­man­ent view of the face of the Earth that con­cerns us with infrared sensors, for example. The lat­ter circle the Earth sev­er­al times a day and provide obser­va­tions along their tra­ject­or­ies with more detailed obser­va­tions. Sci­entif­ic and tech­no­lo­gic­al advances make it pos­sible to dis­tin­guish smal­ler and smal­ler met­eor­o­lo­gic­al objects from space, which is cru­cial for char­ac­ter­ising the state of the atmo­sphere and improv­ing fore­casts of its evolution.

Interview by Isabelle Dumé

Fur­ther reading:

• Mouni­er, A., et al., Detec­tion of Bow Echoes in Kilo­met­er-Scale Fore­casts Using a Con­vo­lu­tion­al Neur­al Net­work, Arti­fi­cial Intel­li­gence for the Earth Sys­tems, 1(2), e210010. https://journals.ametsoc.org/view/journals/aies/1/2/AIES-D-21–0010.1.xml, 2022.
• Rap­port Recher­che 2021 de Météo-France qui fait référence à plusieurs travaux récents au CNRM sur les mod­èles de pré­vi­sion météoro­lo­gique : http://​www​.umr​-cnrm​.fr/​I​M​G​/​p​d​f​/​r​_​r​_​2​0​2​1​-​f​r​_​w​e​b.pdf