Gravity inversion using seismically derived crustal density models and genetic algorithms: An application to the Caribbean-South American Plate boundary

Crustal density models derived from seismic velocity models by means of velocity-density conversions typically reproduce the main features of the observed gravity anomaly over the area but often show significant misfits. Given the uncertainty in the relationship between velocity and density, seismically derived density models should be regarded as an initial estimate of the true subsurface density structure. In this paper, we present a method for estimating the adjustments necessary to a seismically derived density model to improve the fit to gravity data. The method combines the Genetic Algorithm paradigm with linear inversion as a way to approach the non-linear and linear aspects of the problem. The models are divided into three layers representing the sedimentary column, the crystalline crust and the lithospheric mantle; the depths of these layers are determined from the seismic velocity model. Each of the layers is divided into a number of provinces and a density adjustment (Δρ) value is found for each province so that the residual gravity (difference between the observed gravity anomaly and the anomaly calculated for the seismically derived model) is minimized while keeping Δρ between predefined bounds. The preferred position of the province boundaries is found through the artificial evolution of a population of solutions. Given the stochastic nature of the algorithm and the non-uniqueness of the problem, different realizations can yield different solutions. By performing multiple realizations we can analyse a set of solutions by taking their mean and standard deviation, providing not only an estimate of the Δρ distribution in the subsurface but also an estimate of the associated uncertainty. Synthetic tests prove the ability of the algorithm to accurately recover the location of province boundaries and the Δρ values for a known model when using noise-free synthetic data. When noise is added to the data, the algorithm broadly recovers the features that define the known model despite greater standard deviations of the solutions and the occurrence of artefacts in the mean solutions. The algorithm was applied to four profiles across the Caribbean-South America Plate boundary. Some general patterns in the distribution of Δρ were observed consistently in the profiles and are correlated with the interpretations of the velocity models. Positive Δρ values in the sedimentary layer, negative Δρ values for island arc and extended island arc crust with an abrupt change to positive values in South American crust, and positive values in the mantle under the continent and island arc with a transition to negative values under the Caribbean oceanic crust.