Try these ideas

[trevphil]

• ~~Use sampling (GTSAM)~~
• Use tree-reweighted max product
• ~~With LBP, maybe it will converge to the correct marginals if you "observed" N internal bits of SHA-256. Of course, these aren't really observed, but if you brute force all 2^N possibilities and N is small and it converges, it could work~~ [LBP probably won't work]
• ~~Use variational inference (this is not being done by GTSAM DiscreteBayesNet). Can use PyMC3.~~
• Optimization techniques for high-dimensional variables: read here and here
• ~~Particle swarm optimization~~ [too slow, didn't work]
• ~~"Observe" as much as you can from hash outputs, e.g. if the parent is SAME, INV, or AND and the result is 1~~ [done]
• Reformulate AND gate as a set of switchable constraints in which only a subset should be active, and then apply robust estimation techniques like DCS and GNC
• Try to use continuous random variables for 32-bit BitVectors and formulate continuous-domain equivalents for AND, INV, XOR, SHIFT, OR, ADD
• Represent hash function as a neural network (maintaining AND and NOT operations), but incorporate skip connections such that nodes can be made aware of decisions upstream. Look into references at the bottom of this article.