Selection and Justification of Hydro-Morphodynamic Models using Information Theory: Active Learning on Surrogate Emulators

模拟河流生态系统中的水形态动力学过程面临着通过专家驱动的简化假设来重现复杂、动态和高度可变的系统的挑战。因此，在长时空尺度上再现水形态动力学的模型，例如用于气候变化分析的模型，包含巨大的不确定性。模拟水文形态动力学所需的输入数据包括生态系统特征信息，如泥沙粒度和地表高程。然而，每个数据集在时间或空间上都有差距，往往具有相当大的不确定性。因此，建模过程涉及一系列数据采集和处理，以及对复杂系统的大量简化，从而导致各种类型的不确定性。建模链中的这些步骤（及其弱点）构成了复杂的水形态动力学模型的不确定性量化方面的实质性研究挑战。此外，选择多维水形态动力学建模概念是具有挑战性的，因为存在许多不同的建模方法，需要合理的决策。因此，水形态动力学建模可以受益于模型选择、校准和论证的严格和统计方法。为了在可行的计算成本下解决这些建模挑战，我们的项目提出了一种基于贝叶斯分析、信息论和主动学习的机器学习方法，该方法将能够模拟非线性水形态动力学模型。该方法考虑到测量数据的稀疏性，旨在显著缩短计算要求高的模拟时间。解决建模挑战的途径意味着(1)确定性建模的混合建模链的发展；(2)基于随机方法和信息论的代理仿真器；(3)利用模型选择、校准和论证的随机例程；(4)将概念转移到现实世界系统中进行正当性分析。该项目将推动水形态动力学建模从主观确定性工作流发展为复杂的、随机优化的、客观透明的算法序列。

Modelling hydro-morphodynamic processes in river ecosystems faces the challenges to reproduce complex, dynamic, and highly variable systems by making expert-driven simplification hypotheses. For this reason, a model for reproducing hydro-morphodynamics over long spatio-temporal scales, for instance, for climate change analysis, involves vast uncertainty. The input data required for modelling hydro-morphodynamics involve information on ecosystem characteristics, such as sediment grain size and surface elevation. Yet, every dataset has gaps in time or in space with often considerable uncertainty. Thus, the modelling procedure involves a chain of data acquisition and processing, and substantial simplifications of complex systems, which result in various types of uncertainty. These steps (and their weaknesses) in the modelling chain constitute substantial research challenges regarding uncertainty quantification for sophisticated hydro-morphodynamic models. Moreover, the selection of multi-dimensional hydro-morphodynamic modelling concepts is challenging since a multitude of different modelling approaches exist that need justified decisions. Therefore, hydro-morphodynamic modelling can benefit from rigorous and statistical methods for model selection, callibration and justification. To address these modelling challenges at feasible computational costs, our project proposes a machine learning approach based on Bayesian analysis, information theory, and active learning that will enable to emulate non-linear hydro-morphodynamic models. The proposed approach accounts for the sparse nature of measurement data and aims to significantly shorten computationally demanding simulations. The pathway to solving the modelling challenges implies the development of (1) a hybrid modelling chain for deterministic modelling; (2) a surrogate emulator based on stochastic approaches and information theory; (3) stochastic routines to leverage model selection, calibration and justification; and (4) a transfer concept to real-world systems for justifiability analysis. This project will boost hydro-morphodynamic modelling to evolve from a subjective deterministic workflow to a sophisticated, stochastically optimized, and objectively transparent sequence of algorithms.