Support Vector Machines for EM Data Interpretation

Support Vector Machines (SVMs) are a class of trained algorithms which were introduced in the<br>mid-1990s which have rapidly reproduced state-of-the-art computer learning results. They are much<br>easier to understand than neural networks because they mimic the natural way that geophysicists think.<br>Using examples of objects and associated synthetic electromagnetic signatures, an SVM can be trained<br>to find the conductivity of similar objects from new EM signatures. Normally such an operation would<br>fall under the domain of geophysical inversion. SVMs do it in a way that is much more akin to<br>interpolation, producing values for the thickness of a spherical shell based on the similarity of its<br>signature to the training signatures. An SVM can iinterpret a signature with an accuracy of 10% or<br>better. The technique is general enough to apply to a wide variety of geophysical inverse problems.