this is a very nice project I didn’t know about, awesome!

in the reference link you posted about the Gallery of the Street Ratchada it looks pretty clear that the basic XZ grid is respected, which means for that particular facade the corner points of each triangle are only able to slide along the Y axis:

the following detail of the facade section is also very interesting, as it shows that the points of each triangle are not allowed to take “any available Y value”, but they move in fixed steps: https://images.adsttc.com/media/images/589a/3013/e58e/ce85/5d00/000f/slideshow/29_Detail_Facade_Sections_Architectkidd.jpg?1486499852

looks like there are 6 fixed Y positions each corner point of each triangle can take (red line is the base-installation-line, never used, blue lines are the allowed Y positions)

this will become super useful for fabrication, because it will “automatically” cluster the triangular panels of the facade into groups

you have straight triangles and upside-down triangles, but they are the very same ABC triangles (be careful with corner order )

all the triangles have the very same front-projection, so you can easily describe them by just their Y value (again, assuming corner orders is correct):

for instance a triangle [A, B, C] with Y values [3, 3, 4] means points A and B have same Y value 3, and the point C has value 4

to make it easier, you can describe that very triangle also as [0, 0, 1], which is subtracting the smallest Y value for each triangle to all of the Y values of that triangle, because at the end you are just interested in offsets, which is something like “how many Y units is this point offset compared to the Y value of the one with lowest Y in this triangle?”

it becomes easy to see that triangles like [x, x, x+n] are the very same geometrical triangle, so they can be produced in batch, like [x, x, x+n] and [x+n, x+n, x]

not only that, but also other mirroring operation will produce identycal triangles

such as [x, x+n, x] and [x+n, x, x+n] are also the very same triangle, and the following:

**[well, it depends… because if the panels have a different front and a back side (for instance some particular engraving that goes on one side only) then mirroring should not be taken into account, as it “flips” the triangles**

if mirroring is not allowed, then I would also discard clustering triangles based on their surface area, which looked like the fastest and easiest option

generally speaking, the more Y “superpositions” you allow, the more combinations of different triangles you will get

another thing I notice is it looks like this red area has higher offset and deeper valleys, while the bottom left looks more shallow:

you could get similar results using attractors that push Y values more along top right corner, or inside an ellypse-shaped curved with center on top right corner

regarding the base pattern, it looks like diagonal lines runnig bottom right to top left, I believe we could get similar results with image sampling, but could as well be hand modeled (I mean hand-modelling + shallowing attractor) as it’s a very sculptural piece

sorry this is a lot of writing

[edit] some pictures led me to think sometimes a few panels were merged together into bigger ones (like merging 2 or 3 triangles together in some cases, expecially if both of them shared the very same slope) but looking at many more pictures on google made me believe those are all produced as single triangles, installed one by one on the underlying structure, in such a way to leave a beautiful distinctive “wireframe” gap everywhere