Identifying clustering behaviour in EFWI via t-distributed stochastic neighbourhood embedding (t-SNE)

Kristopher A. Innanen, Kevin W. Hall

The dimension reduction and visualization tool known as t-SNE (t-distributed stochastic neighbor embedding) is considered as a possible tool for clustering based regularization of multiparameter elastic FWI. The algorithm is reviewed, and a Matlab implementation (available upon request) is described and illustrated in action on a simple problem. Clustering and regularization are discussed in the context of EFWI, with a recent ``tunneling regularization'' approach being an organizing theme. Simulated initial and final EFWI models are built, and their clustering behaviour is analyzed in the context of the codes. Generally, the use of patterns within low-dimensional t-SNE representations of FWI inputs and outputs as a regularizer, or guide, is plausible, though with most of the big questions still to be broached.