Given a configuration of indistinguishable pebbles on the vertices of a graph, a pebbling move consists of removing two pebbles from one vertex and placing one pebble on an adjacent vertex. The pebbling number of a graph is the least integer such that any configuration with that many pebbles and any target vertex, some sequence of pebbling moves can place a pebble on the target. Graham's conjecture asserts that the pebbling number of the cartesian product of two graphs is at most the product of the two graphs' pebbling numbers. Products of so-called Lemke graphs are widely thought to be the most likely counterexamples to Graham's conjecture, provided one exists. In this paper, we introduce a novel framework for computing pebbling numbers using bilevel optimization. We use this approach to algorithmically show that the pebbling numbers of all products of 8-vertex Lemke graphs are consistent with Graham's conjecture, with the added assumption that pebbles can only be placed on a set of at most four vertices.