17:15 〜 19:15
[MGI27-P01] Manifold learning-aided offline parameter estimation of an Earth system model using observation of changing climate
キーワード：Uniform Manifold Approximation and Projection、Surrogate-model-based Parameter estimation、Earth system model of intermediate complexity
Earth system model is a powerful tool to predict future climate change due to anthropogenic global warming. Earth system models have achieved realistic representations of the Earth but still have significant uncertainty in reliably predicting future climate change. Offline parameter estimation method is one of the most effective ways to reduce the model uncertainty by utilizing observation. In this work, we explore a novel method to quantify and reduce the parametric uncertainty of Earth system models and evaluate its effect on the accuracy of model-based prediction.
In previous studies, there are mainly two ways to reduce the parametric uncertainty: Surrogate-model-based Parameter estimation (SP) (Yarger et al., 2023) and Ensemble Kalman Inversion (EKI) (Kovachki and Stuart, 2018). SP is a non-parametric uncertainty quantification and reduction method and can globally explore the parameters. Since SP needs to average climate time series data, it has to neglect potentially meaningful variability of the observation data. On the other hand, EKI can directly use time-series data. However, EKI needs the assumption that the posterior distribution of model parameters follows Gaussian distribution and the relationship between parameters and model-observation fitting is quasi-linear.
In this work, we combined Surrogate-model-based Parameter estimation techniques with Uniform Manifold Approximation and Projection (UMAP), a manifold learning method for dimensionality reduction, and developed the Surrogate-model-based Parameter Estimation method with UMAP (UMAP-SP). UMAP-SP can effectively reduce parameter uncertainty even when directly applied to high-dimensional climate time series data. This makes UMAP-SP a valuable tool for analyzing observations of climate change due to anthropogenic global warming, which will lead to successful prediction of future climate change including climate tipping. We tested this method on an Earth system model of intermediate complexity, LOVECLIM (Goose et al., 2010), to enhance model prediction accuracy. We found that UMAP can effectively reduce the dimension of simulated data preserving their important structure and led to a successful estimation of the posterior distribution of model parameters.
In previous studies, there are mainly two ways to reduce the parametric uncertainty: Surrogate-model-based Parameter estimation (SP) (Yarger et al., 2023) and Ensemble Kalman Inversion (EKI) (Kovachki and Stuart, 2018). SP is a non-parametric uncertainty quantification and reduction method and can globally explore the parameters. Since SP needs to average climate time series data, it has to neglect potentially meaningful variability of the observation data. On the other hand, EKI can directly use time-series data. However, EKI needs the assumption that the posterior distribution of model parameters follows Gaussian distribution and the relationship between parameters and model-observation fitting is quasi-linear.
In this work, we combined Surrogate-model-based Parameter estimation techniques with Uniform Manifold Approximation and Projection (UMAP), a manifold learning method for dimensionality reduction, and developed the Surrogate-model-based Parameter Estimation method with UMAP (UMAP-SP). UMAP-SP can effectively reduce parameter uncertainty even when directly applied to high-dimensional climate time series data. This makes UMAP-SP a valuable tool for analyzing observations of climate change due to anthropogenic global warming, which will lead to successful prediction of future climate change including climate tipping. We tested this method on an Earth system model of intermediate complexity, LOVECLIM (Goose et al., 2010), to enhance model prediction accuracy. We found that UMAP can effectively reduce the dimension of simulated data preserving their important structure and led to a successful estimation of the posterior distribution of model parameters.