As you mention in a comment, the four layers in the Y-dimension have a 20%, 40%, 60% and 80% probability of true values. This means that e.g. if you look at a 5x4x5 block, you would expect its four layers to have around 5, 10, 15 and 20 true values. This knowledge of the probability in each layer can be used to speed up the search.

Imagine that you're looking for a pattern in a 5x4x5 block, and the number of true values in its four layers are 8, 8, 20 and 16. This means that the difference between the actual number of true values and the expected number is 3, 2, 5 and 4.
It would be beneficial if you searched for the pattern by looking at the layer with difference 5 first, then 4, then 3 and then 2, because the 1-layer pattern with the highest difference from the expected number of true values is likely to be less common, meaning that you can discard more locations without having to look at different layers before finding a difference.

You could expand this idea by looking at how the true values are distributed per layer. If in the example above, the pattern had 5 true values in the first layer, that would be exactly as expected. However, if in the 5x5 grid these 5 true values formed a line or a diagonal, this may be an unusual pattern in your data, and this would make this layer a candidate to be checked first.

I wouldn't expect the advantage of this method to be huge, but it should be noticeable, and it does use the knowledge of the probability of true values in the different layers.