Modellunsicherheiten in Klimafolgeprojektionen. Eine Einführung
Abstract
Um Folgen des Klimawandels abzuschätzen, erzeugen Klimaforscher oft Modellvorhersagen basierend auf einem Emissionsszenario, sogenannte Klimafolgeprojektionen. Solche Modellprojektionen sind mit Unsicherheiten behaftet, die sich aus den Unsicherheiten der einzelnen Modellkomponenten (z.B. Treibern, Parametern oder Struktur) ergeben. Die Quantifizierung dieser Unsicherheiten ist für die Interpretation der Projektionen durch Forschung, Entscheidungsträger und Öffentlichkeit von essenzieller Wichtigkeit. Ziel dieses Artikels ist es, eine Einführung in die Berechnung von Modellunsicherheiten im Kontext von Klimafolgeprojektionen zu geben. Wir diskutieren hierzu zunächst die Gründe und Typen von Modellunsicherheiten und erklären danach wie man diese quantifizieren (=Unsicherheitsanalyse) und über den Abgleich mit Beobachtungsdaten reduzieren (=Modellinversion) kann. Wir veranschaulichen die besprochenen Konzepte durch eine Fallstudie, in der wir eine Unsicherheitsanalyse des terrestrischen Ökosystemmodells LPJ-GUESS für einen Standort in Franken vornehmen.
What the paper shows and why it matters (AI-generated)
Researchers estimate the consequences of climate change with climate impact projections: simulations of how a system such as the climate or the vegetation develops under an assumed scenario, for example unchecked CO₂ emissions. These are conditional “if–then” calculations, not forecasts with known probabilities, and they come from mathematical models because the systems are too complex for intuition. Such models rest on inputs, parameters and process assumptions that cannot all be determined exactly, so their results are uncertain, and researchers, decision-makers and the public need to know how uncertain to interpret them. The paper sets out to give an accessible introduction to where this uncertainty comes from and how it can be quantified.
The paper argues that quantifying the uncertainty of predictions of complex system models, such as projections of climate impacts, is essential for interpreting them, and explains in introductory form how this is done. It traces uncertainty to the starting state, the inputs, the parameters, the model structure and tipping points (critical thresholds beyond which a system changes qualitatively), and distinguishes risk (probabilities of possible futures known), uncertainty (futures known, probabilities not) and deep uncertainty (not even all futures known). To quantify it, the paper presents sensitivity analysis (how strongly the output reacts to a change in an input, parameter or other model assumption), Monte Carlo propagation (running the model with many randomly drawn input combinations and examining the spread of results) and model inversion (adjusting the model against observations, which narrows the uncertainty). In a worked example, the study varies the parameters of the vegetation model LPJ-GUESS by an assumed ±10% for pedunculate oak in one grid cell in Franconia. Under present-day climate, deviations from the reference run towards lower aboveground carbon production come out clearly more plausible than deviations towards higher production, and the sensitivity spreads over more parameters.
The paper presents these techniques as generally applicable, with particular relevance for climate change, and treats the oak simulation as an illustration, calling its uncertainties fairly large because the parameter uncertainty was assumed to be large and only one grid cell was examined. It advises model inversion despite long run times, and notes that structural uncertainty and tipping points are often hard to quantify probabilistically and that communicating uncertainty is a challenge in its own right.