Constraining the Large Uncertainties in Earth System Model Projections with a Big Data Approach

Økende bevissthet om menneskeskapte klimaendringer har ført til en eksponsiell økning i mengden av klimadata fra både observasjoner og modeller de siste tiårene. Dette introduserer nye utfordringer for forskningsmiljøet med tanke på hvordan man effektivt kan analysere og tolke slike store mengder data. Prosjektet COLUMBIA (Constraining the Large Uncertainties in Earth System Model Projections with Big Data Approach) har som mål å adressere dette ved å ta i bruk avanserte statistiske metoder og nye maskinlæringsverktøy for å analysere klimamodellprojeksjoner. Det tverrfaglige teamet, som består av forskere innenfor natur- og beregningsvitenskapen, har blant annet brukt maksinlæringsmetoder til å gi mer nøyaktige estimater av karbonsluk i havet. Forskningen viste at unøyaktigheter i modellenes representasjon av blanding i Nord-Atlanteren og Sørishavet førte til divergens i de simulerte karbonfluksene. Golfstrømmen og den subantarktiske fronten ble identifisert som nøkkelregioner hvor fremtidige observasjoner vil være avgjørende for å forbedre vår forståelse av hvordan havets karbonsluk vil reagere på fremtidige CO2-utslipp. Vi utviklet en ny klyngemetode som kan hjelpe oss å evaluere i hvilken grad klimamodeller er i stand til å simulere en realistisk ENSO (El Nino Southern Oscillation), den viktigste modusen for global klimavariabilitet. Mange modeller har fortsatt problemer med å reprodusere de observerte klimatiske egenskapene i det tropiske Stillehavet til tross for en markant økning i både kompleksitet og mengden av beregningsressurser som benyttes. Analysen basert på vår metode indikerer at fremtidige klimahendelser knyttet til ENSO, slik som tørke og skogbrann, vil bli hyppigere og øke i styrke. Fortsatt miljøovervåking er avgjørende for å forbedre vår forståelse av klimasystemet og for å bedre kunne forutsi fremtidige endringer. På grunn av de enorme havområdene på jorden og de relativt begrensede forskningsressurser som er tilgjengelig er det umulig å observere alle relevante variabler i havet globalt. Vi bruker derfor maskinlæring sammen med klimamodelldata til å identifisere viktige havregioner der de fysiske og biogeokjemiske egenskapene i stor grad vil bli påvirket av fremtidige klimaendringer: Sørishavet, det tropiske Stillehavet og Nord-Atlanteren. Langsiktige klimaovervåkingsnettverk i disse nøkkelområdene vil kunne gi verdifull kunnskap for å innskrenke mulige fremtidige klimaprojeksjoner.

The project has produced a number of key results in the field of ocean carbon cycle and future climate projections. These findings have been published in high-level peer-reviewed journals (25 scientific publication at the time of reporting with more in the pipeline). In addition, the team has been active in disseminating results in key conferences, meetings, and workshops (40 presentations and more scheduled beyond the project period). The project recruited five early-career researchers, who have all gained research training in relatively novel field, combining climate and computational sciences. Each of the young researcher have been involved or led at least one publication associated with the project results. As they are well integrated to the Bjerknes Centre for Climate Research in Bergen, they will have the opportunity to apply their novel skills on broader field of climate change research after the project ends. An example is postdoc Timothee Bourgeois, who gained new skills in analyzing high volume of Earth system model outputs. His new competence allows him to be hired permanently at NORCE (as Forsker II) and to participate and contribute to the ongoing H2020-COMFORT and H2020-OceanNETs projects. Some senior members of the project teams have been invited to present key project findings at high-level meetings. A few examples are: (i) Science Day: AI in the fight against climate change, organized by the British Embassy in Oslo (Johannsen), (ii) invited to join the European Marine Board Forum working group on Marine Big Data (Tjiputra), (iii) invited to a Bilateral Artificial Intelligence Forum, join-organized by the Norwegian Ministry of Education and Research and the US Department of Energy (Tjiputra), and (iv) invited to lead/co-lead synthesis tasks in model evaluation (Goris), North Atlantic carbon sink (Tjiputra), and seasonal cycle of ocean CO2 fluxes (Schwinger) of the ongoing international RECCAP2 (Regional Carbon Cycle Assessment and Processes) synthesis project. The project activities have established new collaborations and opportunities to accelerate progress in both natural and computational fields. Several initiatives are in the pipeline to pursue more join research beyond COLUMBIA. Our most significant contribution to the field is by better elucidating the source of model uncertainty in projecting future ocean carbon sinks and other climate projections. These have provided essential new knowledge for both the modeling and observing communities, i.e. (i) guidance on how to focus future Earth system model development, in preparation for the IPCC-AR7 and (ii) strategy for future marine monitoring network to better understand the consequences of anthropogenic climate change. Moving forward, the project team are invited to contribute to the newly funded Horizon-Europe project (OceanICU), which will use the knowledge gained from COLUMBIA to better constrain future ocean carbon sink associated with biological activity.

The increasing volume of climate model data makes the use of traditional analysis tool impractical. This hinders the discovery of new crucial knowledge for society. COLUMBIA initiates a multidisciplinary research activity that will result in a cutting-edge customizable methodology that aims to facilitate rigorous evaluation of large ensemble of Earth system models. The innovative assessment tool will be based on combining Machine Learning with robust statistics and multiple process-oriented assessments, and aims to filter 'good' models from the rest and therefore increasing our confidence when synthesizing future climate change projections. The proposed work is well timed as the process for upcoming IPCC-AR6 has just began and the first batch of new Earth system model simulations will become available in early 2018. The proposed integrated evaluation method is at the forefront of international efforts in understanding sources of model uncertainty and producing optimized future climate projections. The project outcome will therefore have a broad international relevance for both climate modeling and observational community. Customizing existing Machine Learning techniques for climate science will also brings new opportunities and values to the computer science community. New scientific knowledge gained from the project will provide critical support for the climate science community as well important inputs for the IPCC-AR6. A unique close collaboration between climate modellers and computational scientists exists within Uni Research and the Bjerknes Centre, ensuring seamless collaboration in developing the new method. Ultimately, method developed in COLUMBIA will produce a new generation of optimized climate change projections that will give the best advice to policy makers and help society to find the best adaption strategy for climate change.