Seminar: Mauricio Lima

Land and atmosphere interact through fluxes of water, energy and carbon. The relations between variables in this coupled system are not only nonlinear and uncertain, but uncertainties are also nonlinear themselves. We propose a framework that exploits these properties through generative diffusion models, which can be used to elucidate and quantify ecosystem-weather interactions from one single learned prior. We demonstrate the framework through two experiments. First, we show that, while forests become more resilient under warmer temperatures, grasslands become more likely to fail in the same conditions, even though the mean response of net carbon flux is similar for both ecosystems. This shows that extreme events can diverge from mean responses. Second, we show that, while air temperature controls soil moisture through vapour pressure deficit in forests, savannas and shrublands are more sensitive to both soil moisture and vapour pressure deficit in the first place. This shows that the covariance matrix from the coupling contains fundamental information from interactions. Our results motivate further studies making use of the joint probability distribution of variables in the land-atmosphere coupled system, rather than solely considering mean responses or quantifying uncertainties simply with standard deviations.