Stability criteria for Bayesian calibration of reservoir sedimentation models

Abstract

Modeling reservoir sedimentation is particularly challenging due to the simultaneous simulation of shallow shores, tributary deltas, and deep waters. The shallow upstream parts of reservoirs, where deltaic avulsion and erosion processes occur, compete with the validity of modeling assumptions used to simulate the deposition of fine sediments in deep waters. We investigate how complex numerical models can be calibrated to accurately predict reservoir sedimentation in the presence of competing model simplifications and identify the importance of calibration parameters for prioritization in measurement campaigns. This study applies Bayesian calibration, a supervised learning technique using surrogate-assisted Bayesian inversion with a Gaussian Process Emulator to calibrate a two-dimensional (2d) hydro-morphodynamic model for simulating sedimentation processes in a reservoir in Albania. Four calibration parameters were fitted to obtain the statistically best possible simulation of bed level changes between 2016 and 2019 through two differently constraining data scenarios. One scenario included measurements from the entire upstream half of the reservoir. Another scenario only included measurements in the geospatially valid range of the numerical model. Model accuracy parameters, Bayesian model evidence, and the variability of the four calibration parameters indicate that Bayesian calibration only converges toward physically meaningful parameter combinations when the calibration nodes are in the valid range of the numerical model. The Bayesian approach also allowed for a comparison of multiple parameters and found that the dry bulk density of the deposited sediments is the most important factor for calibration.

Bayesian Calibration Points to Misconceptions in Three‐Dimensional Hydrodynamic Reservoir Modeling

“…The problem of too small (too close to zero) probabilities in the framework of active learning is also referred to as curse of dimensionality (Bellman, 1957), which is discussed in detail by Mouris et al. (2023).…”

“…Because the total of posterior probabilities still needed to sum up to 1, the additional dimension would decrease the prior output space to infinitesimally small numbers, which would all be rejected in the rejection sampling step (Oladyshkin et al, 2020;Smith & Gelfand, 1992). The problem of too small (too close to zero) probabilities in the framework of active learning is also referred to as curse of dimensionality (Bellman, 1957), which is discussed in detail by Mouris et al (2023).…”

Information‐Theoretic Scores for Bayesian Model Selection and Similarity Analysis: Concept and Application to a Groundwater Problem

“…For brevity, we will call this multi‐fidelity model comparisons. In said cases, simplifications could require averaging available observations or ignoring subsets and/or types of data (e.g., see Mouris et al., 2023). Multi‐fidelity comparisons can be useful, not only to select the best model, but also to quantify the change in model performance under different model configurations.…”

“…The paper, however, does not present a specific application. The approach proposed in Oladyshkin and Nowak (2019) has been applied in active learning techniques for surrogate model generation, which closely resembles optimal experimental design setups (Mouris et al., 2023; Oladyshkin et al., 2020), but not, to the authors' knowledge, for model selection or similarity analysis.…”