ALBERT

Description: 〈span〉〈div〉Summary〈/div〉Non-uniqueness in the geophysical inverse problem is well recognized and so too is the ability to obtain solutions with different character by altering the form of the regularization function. Of particular note is the use of ℓ〈sub〉〈span〉p〈/span〉〈/sub〉-norms with 〈span〉p〈/span〉 ∈ [0, 2] which gives rise to sparse or smooth models. Most algorithms are designed to implement a single ℓ〈sub〉〈span〉p〈/span〉〈/sub〉-norm for the entire model domain. This is not adequate when the fundamental character of the model changes throughout the volume of interest. In such cases we require a generalized regularization function where each sub-volume of the model domain has penalties on smallness and roughness and its own suite of ℓ〈sub〉〈span〉p〈/span〉〈/sub〉 parameters. Solving the inverse problem using mixed ℓ〈sub〉〈span〉p〈/span〉〈/sub〉-norms in the regularization (especially for 〈span〉p〈/span〉 〈 1) is computationally challenging. We use the Lawson formulation for the ℓ〈sub〉〈span〉p〈/span〉〈/sub〉-norm and solve the optimization problem with Iterative Reweighted Least Squares. The algorithm has two stages; we first solve the 〈span〉l〈/span〉〈sub〉2〈/sub〉-norm problem and then we switch to the desired suite of ℓ〈sub〉〈span〉p〈/span〉〈/sub〉-norms; there is one value of 〈span〉p〈/span〉 for each term in the objective function. To handle the large changes in numerical values of the regularization function when 〈span〉p〈/span〉-values are changed, and to ensure that each component of the regularization is contributing to the final solution, we successively rescale the gradients in our Gauss-Newton solution. An indicator function allows us to evaluate our success in finding a solution in which components of the objective function have been equally influential. We use our algorithm to generate an ensemble of solutions with mixed ℓ〈sub〉〈span〉p〈/span〉〈/sub〉-norms. This illuminates some of the non-uniqueness in the inverse problem and helps prevent over-interpretation that can occur by having only one solution. In addition, we use this ensemble to estimate the suite of 〈span〉p〈/span〉-values that can be used in a final inversion. First, the most common features of our ensemble are extracted using principal component analysis and edge detection procedures; this provides a reference model. A correlation of each member of the ensemble with the reference model, carried out in a windowed domain, then yields a set of 〈span〉p〈/span〉-values for each model cell. The efficacy of our technique is illustrated on a synthetic 2D cross-well example. We then apply our technique to the field example that motivated this research, the 3D inversion of magnetic data at a kimberlite site in Canada. Since the final regularization terms have different sets of 〈span〉p〈/span〉-values in different regions of model space we are able to recover compact regions associated with the kimberlite intrusions, continuous linear features with sharp edges that are associated with dykes, and a background that is relatively smooth. The result has a geologic character that would not have been achievable without the use of spatially variable mixed norms.〈/span〉