Abstract: At present, local volume spline is one of main modeling methods in the visualization of large scattered volumetric data sets. A good partition on the modeling domain is crucial to the implementation of local volume spline. A simulated annealing algorithm is presented to carry out optimized partition of the domain; two experiments together with comparison with Nielson's results are described to indicate the feasibility of this approach and some limits to the size of data sets. Finally, as an example, some visual images for a set of oceanic investigation data are given to show the potential of this method.