High-dimensional parameter optimization of a biogeochemical model: a multi-variable BGC-Argo data assimilation approach

Engineering notes

Key topics: autonomous driving, control. See the paper for implementation details and experimental results.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Abstract. The predictive accuracy of marine biogeochemical models is limited by uncertainty in their parameter values. We present a parameter optimization framework using iterative Importance Sampling (iIS) to constrain the PISCES biogeochemical model within a one-dimensional representation of the ocean by leveraging the comprehensive, multi-variable dataset provided by Biogeochemical-Argo (BGC-Argo) floats. Using seasonal observations from a single BGC-Argo float in the North Atlantic during the year 2015, we assimilate twenty observational quantities derived from eight biogeochemical tracers to directly optimise all 95 poorly known model parameters. A prerequisite global sensitivity analysis (GSA) identifies parameters controlling zooplankton dynamics as the dominant source of model sensitivity at this site. We compare three strategies: (1) optimising a subset of parameters selected for their strong direct influence (Main effects); (2) optimising a larger subset that also includes parameters influential through non-linear interactions (Total effects); and (3) simultaneously optimising all 95 parameters. All three approaches achieve a statistically indistinguishable goodness of fit to the assimilated observations, reducing the median normalised RMSE across metrics by 54 %–56 % relative to the reference simulation with default PISCES parameters. The comprehensive, multi-variable observational constraint yields posterior parameter distributions with negligible inter-parameter correlation, shifting the long-standing challenge of correlated equifinality to uncorrelated equifinality: multiple optimal parameter sets exist, but the individual parameters are uncorrelated rather than compensating for one another. Parameter uncertainty is reduced by 16 %–41 % relative to the broad uniform prior distributions. While all strategies produce a similar goodness of fit for the assimilated variables, they differ in computational cost and in their estimation of uncertainty for unassimilated variables. The All-parameters strategy provides a fuller accounting of parametric uncertainty for unassimilated variables, because it allows all parameters to vary rather than artificially fixing poorly known parameters at their default values. The method is computationally tractable thanks to the use of a one-dimensional model configuration, requiring approximately 24 CPU-hours for the optimization step, whereas the prerequisite GSA was ∼ 40 times more computationally expensive. When implemented in the three-dimensional IBI (Iberian–Biscay–Irish) regional model at 1/36° resolution over a three-year period (2017–2019), the optimized parameter set from the All-parameters strategy improves the overall normalised RMSE against approximately 1430 independent BGC-Argo vertical profiles of nutrients and carbonates, although nitrate and silicate show degraded skill in the Mediterranean. Surface chlorophyll a is also better reproduced across the domain, when evaluated against a multi-observation reprocessed product. The generality of the approach should be tested by applying it to additional BGC-Argo floats in other oceanic regions.

Links and sources

• PDF from original source

Comments