Surrogate-based optimization using an artificial neural network for a parameter identification in a 3D marine ecosystem model

<strong>Abstract:</strong> Parameter identification for marine ecosystem models is important for the assessment and validation of marine ecosystem models against observational data. The surrogate-based optimization (SBO) is a computationally efficient method to optimize complex models. SBO replaces the computationally expensive (high-fidelity) model by a surrogate constructed from a less accurate but computationally cheaper (low-fidelity) model in combination with an appropriate correction approach, which improves the accuracy of the low-fidelity model. To construct a computationally cheap low-fidelity model, we tested three different approaches to compute an approximation of the annually periodic solution (i.e., a steady annual cycle) of a marine ecosystem model: firstly, a reduced number of spin-up iterations (several decades instead of millennia), secondly, an artificial neural network (ANN) approximating the steady annual cycle and, finally, a combination of the both approaches. Except for the low-fidelity model using only the ANN, the SBO yielded a solution close to the target and reduced the computational effort significantly. If an ANN approximating appropriately a marine ecosystem model is available, the SBO using this ANN as low-fidelity model presents a promising and computational efficient method for the validation. <strong>Content:</strong> SQLite database including the data of the different optimization runs Structure and weights of the used artificial neural network Tracer concentrations obtain from the high-fidelity model for the different optimization runs摘要：海洋生态系统模型的参数辨识，对于基于观测数据开展海洋生态系统模型的评估与验证工作至关重要。基于代理的优化（surrogate-based optimization, SBO）是一种计算高效的复杂模型优化方法，其通过结合精度较低但计算成本更低的低精度（low-fidelity）模型与合适校正手段构建的代理模型，替代计算开销高昂的高精度（high-fidelity）模型，以此提升低精度模型的精度。为构建计算成本低廉的低精度模型，我们测试了三种用于近似海洋生态系统模型年度周期解（即稳定年度循环）的方法：其一为减少自旋迭代次数（采用数十年而非数千年的迭代时长）；其二为利用人工神经网络（artificial neural network, ANN）近似稳定年度循环；其三为上述两种方法的组合。除仅使用人工神经网络的低精度模型外，其余基于代理的优化方案均得到了接近目标值的解，并显著降低了计算开销。若可获取能够适当近似海洋生态系统模型的人工神经网络，以该人工神经网络作为低精度模型的基于代理的优化方法，将是一种极具应用前景且计算高效的模型验证手段。 数据集内容：包含各类优化运行数据的SQLite数据库；所用人工神经网络的结构与权重；各类优化运行对应的高精度模型模拟得到的示踪剂浓度数据。