Selecting a global optimization method to estimate the oceanic particle cycling rate constants

The objective is to select an inverse method to estimate the parameters of a dynamical model of the oceanic particle cycling from in situ data. Estima ting the parameters of a dynamical model is a nonlinear inverse problem, ev en in the case of linear dynamics. Generally, biogeochemical models are cha racterized by complex nonlinear dynamics and by a high sensitivity to their parameters. This makes the parameter estimation problem strongly nonlinear . We show that an approach based on a linearization around an a priori solu tion and on a gradient descent method is not appropriate given the complexi ty of the related cost functions and our poor a priori knowledge of the par ameters. Global Optimization Algorithms (GOAs) appear as better candidates. We present a comparison of a deterministic (TRUST), and two stochastic (si mulated annealing and genetic algorithm) GOAs. From an exact model integrat ion, a synthetic data set is generated which mimics the space-time sampling of a reference campaign. Simulated optimizations of two to the eight model parameters are performed. The parameter realistic ranges of values are the only available a priori' information. The results and the behavior of the GOAs are analyzed in details. The three GOAs can recover at least two param eters. However, the gradient requirement of deterministic methods proves a serious drawback. Moreover, the complexity of the TRUST makes the estimatio n of more than two parameters hardly conceivable. The genetic algorithm qui ckly converges toward the eight parameter solution, whereas the simulated a nnealing is trapped by a local minimum. Generally, the genetic algorithm is less computationally expensive, swifter to converge, and has more robust p rocedural parameters than the simulated annealing.