SAMFURORIN YANAYI NA AI SUN FI BIN DIDDIGIN GUGUWA, AMMA SUNA RAUNANA SU FIYE DA KIMA

Samfurorin yanayi da aka horar da su a kan bayanai yanzu suna fafatawa da mafi kyawun tsarin hasashe masu tushe a kimiyyar lissafi a matsakaici, da ɗan ƙaramin kaso na kuɗin kwamfuta. Ko hakan yana tabbata ga abubuwa masu tsanani da ba kasafai ake gani ba — waɗanda a cikinsu kyakkyawan hasashe ya fi muhimmanci — ana jayayya a kai. Binciken da aka yi a baya ya gano manyan gazawa a kan yanayi mai karya tarihi da guguwa masu zafi masu ƙarfi, amma galibi sun gwada samfurori masu hasashe ɗaya tak (deterministic), waɗanda ke ba da hasashe guda ɗaya. Sabbin samfurori kuma na yiwuwa ne (probabilistic): suna samar da rukunin makomai masu yiwuwa (ensemble), kowanne “memba”.

Guguwa masu zafi (tropical cyclones) gwaji ne mai tsauri. Suna cikin haɗurran yanayi da suka fi ɓarna, kuma kwashe mutane ya dogara ne da hasashen wurinsu da ƙarfinsu kwanaki da yawa kafin lokaci, tare da ma’auni na gaskiya na rashin tabbas. Duk da haka suna mamaye kusan kashi 0.5% ne kawai na wuraren grid a cikin kwaikwayon yanayin duniya.

Guguwa 178, an sake kwaikwayon su

Nathan Chalumeau da abokan aikinsa — daga INRIA, Jami’ar Sorbonne da CNRS, da AXA a Paris — sun kwatanta samfurorin AI uku: GenCast, wanda ke zaɓar yanayin gaba ta hanyar samfurin yaɗuwa (diffusion model); FourCastNet 3, wanda NVIDIA ta fitar; da FGN, wanda ke gina rukuninsa daga cibiyoyin jijiyoyi huɗu da aka horar daban-daban. Ma’aunin su shi ne IFS-ENS, rukunin hasashe na aiki na ECMWF mai tushe a kimiyyar lissafi, wanda aka yi amfani da shi a matsayin tushen kwatanci na lissafi.

Sun zaɓi guguwa 178 waɗanda suka kai aƙalla ƙulli (knots) 64 na iska mai ɗorewa, a kowane yankin teku, a cikin yanayi biyu: guguwa 125 daga 2021–2023 tare da samfuran da aka horar a kan sake nazarin ERA5, da guguwa 53 daga 2025 tare da sigogin aiki da aka gyara a kan nazarce-nazarcen ECMWF masu inganci sosai. Kowane hasashe ya yi kwanaki 15 da membobi 49 ko 50, yana farawa awanni 12 kafin rahoton iska na farko na guguwar — hasashen membobi 26,541 gaba ɗaya. Sake gudanar da hasashen 2021–2023 ya ɗauki fiye da awanni 6,000 na na’urorin sarrafa hotuna na H100.

Auna gajimaren wurare masu yiwuwa

Yawanci ana auna hanyoyin hasashe da matsakaicin wurin rukunin. Marubutan sun nuna dalilin da ya sa hakan zai iya ɓatar da mutum da misalin Guguwar Typhoon Bualoi a 2025: membobin IFS-ENS sun rabu zuwa gungu daban-daban waɗanda babu wani matsakaici da zai iya wakiltar su.

Typhoon Bualoi (2025): wuraren da FGN (ja) da IFS-ENS (shuɗi) suka yi hasashe a lokuta huɗu na gaba, kewaye da hanyar da aka gani (ɗigo baƙaƙe). — Hoto na 2, Chalumeau et al. (2026), arXiv:2610.22077.

Maimakon haka, sun mayar da kowane rukuni zuwa taswirar yiwuwa suka tambayi nawa daga cikinta ya faɗa a kan wurin da aka ga guguwar. Ƙananan maki na nufin hasashe mafi kyau.

A kan guguwar 2021–2023, GenCast ya doke IFS-ENS a kan kashi 68% na guguwa, FourCastNet 3 kuma a kan kashi 58% — ƙananan rata. A kan guguwar 2025, ratar ta faɗaɗa: FGN ya doke IFS-ENS a kan kusan kashi 90% na guguwa 41 da aka riƙe, kuma ya zo na farko cikin samfurori uku a kan kashi 59% nasu, idan aka kwatanta da kashi 34% na GenCast da kashi 7% na IFS-ENS. FGN kuma ya sami maki mafi kyau a kan cikakkun filayen yanayi, kuma duka samfurorin AI biyu sun fi samun ci gaba a kan IFS-ENS kewaye da guguwa fiye da a duk faɗin duniya. Lokacin da aka kwatanta yankunan iska mafi ƙarfi da matsin iska mafi ƙanƙanta bisa siffarsu da wurinsu kaɗai, tsarin ya kasance iri ɗaya: FGN, sannan GenCast, sannan IFS-ENS.

Guguwa masu laushi fiye da kima

Sanya wuri mafi kyau bai nufin ƙarfi mafi kyau ba. A kan guguwar 2025, FGN da GenCast duka sun rage ƙimar iskoki mafi ƙarfi a gajeren lokaci da matsakaicin lokaci, yayin da IFS-ENS ya kasance cikin kashi 10% na ma’aunin. Gyara a kan bayanan aiki bai kawar da wannan karkata ba. A yanayin farko, GenCast ya sami maki na gida mafi kyau amma kuma ƙarancin iska mafi girma. Rukunonin kuma ba a daidaita su da kyau ba, da ƙaramin bazuwa a wasu wurare.

Akwai wani juyi. Guguwar AI sun fi rauni, amma alaƙar da ke tsakanin iska da matsin iska a cikinsu tana da kyau. Da zarar an cire karkatar ƙarfi ta kowane samfuri ta hanyar jera su bisa matsayi, alaƙar da ke tsakanin iska mafi girma da matsin iska mafi ƙanƙanta ita ce −0.83 ga FGN da −0.84 ga GenCast, idan aka kwatanta da −0.88 ga nazarin mai inganci sosai da −0.87 ga IFS-ENS. Ga FGN, wanda aka horar da shi a kan maɓalli ɗaya da wuri ɗaya a lokaci guda, ba a taɓa tilasta wannan alaƙar ba. Duk da haka ta bayyana.

Hagu: guguwar samfurorin AI suna da iskoki masu rauni da ramukan matsin iska marasa zurfi fiye da ma’aunin. Dama: da zarar an jera su, kowane tushe yana bin kusan alaƙa ɗaya tsakanin iska da matsin iska. — Hoto na 6, Chalumeau et al. (2026), arXiv:2610.22077.

Ba tsarin gargaɗi ba — tukuna

Marubutan da kansu sun shata iyaka ga sakamakonsu. Binciken yana sake kwaikwayon zaɓaɓɓun guguwa na baya; bai haɗa da lokutan da babu guguwa ba, don haka ba zai iya yanke hukunci ko samfurorin suna ba da gargaɗin ƙarya ba. Yanayin biyu sun bambanta a guguwa, sigogin samfuri har ma da haɓaka IFS-ENS, don haka ba za a iya ware tasirin gyaran ba. Kuma bayanin ɗabi’a a fili yake: wannan sakamakon “bai kamata a karanta shi a matsayin jagora kan amfani da ɗaya daga cikin waɗannan samfurorin don gargaɗin aiki ba.”

Ƙarshensu jerin ayyuka ne. A gyara ƙarfin, a tsara manufofin horo da ke ba iskoki masu tsanani nauyi, a inganta rashin tabbas — kuma a duba kowane lokaci cewa ci gaban da aka samu a wuri da kuma a alaƙar iska da matsin iska ya tsira.