McNeil, Ben I., Andrew R. Jacobson, Robert M. Key and Jorge L. Sarmiento

AOS Program, Forrestal Campus, Princeton University, NJ, 0854, USA, Tel: 609-258-0979, E-mail: bmcneil@princeton.edu

On the use of global optimization schemes for water-mass analysis and remineralization ratios

Classic water-mass analysis known as Optimum Multiple Parameteric (OMP) analysis has been used to quantify large scale mixing processes in the ocean. OMP has also recently been extended to quantify remineralization ratios and help reconstruct anthropogenic CO₂ in the ocean. The inherent assumption of OMP is that conservative properties (Temp, Sal) and non-conservative biogeochemical properties in the ocean (Nit, Phos, DIC etc.) can be solved in a linear way. Anderson and Sarmiento (1994) used a local non-linear optimization scheme to determine water mass mixing and remineralization ratios. We explore the validity of 3 different optimization schemes in determining water-mass mixing and remineralization ratios in the ocean. To do this we ran a Monte-carlo based forward model through our objective function (setup equations) to construct “synthetic” observations. Many different synthetic data-sets were constructed that included all realistic types of mixing (2-5 endmembers) and the inherent uncertainties in the measurements. We then use three different optimization schemes (OMP and two global optimization schemes known as Simulated Annealing-SA and Genetic Algorithm-GA) to solve the set of equations (i.e., find the minimum of the objective function) and determine how well each scheme solves the known parameters (i.e., water-mass fractions and remineralization ratios). SA and GA are both globally stochastic non-linear schemes which have the ability to ‘jump’ out of local minima to find the global minimum of the objective function. Based on our ‘synthetic’ results, we find that the OMP scheme fails to find the correct solution for all different scenarios. The reason for this is due to non-linearity introduced by the non-conservative parts of the model setup. Due to this non-linearity, we find that the global optimization schemes to be far superior to OMP in solving for both mixing and remineralization ratios in the ocean. Of the global optimization schemes, we find Simulated Annealing to be the most efficient and robust at finding the correct solution even in a relatively complex mixing regime. We present results of both water mass analysis and remineralization ratios for the Indian Ocean by applying Simulated Annealing to the WOCE data set.