Human aesthetics obstructs tech efficiency
Letting computers work out design often results in incomprehensible forms – and that’s exactly what’s required. Richard A Lovett reports.
Computer-assisted advances in optics, photonics, and related sciences are producing new designs so different from anything humans would develop on their own that at first glance they can seem almost incomprehensible.
“These structures are non-intuitive,” Boaz Blankrot, a Ph.D. student in applied mathematics at Vienna University of Technology, Austria, recently told a meeting of the Society for Industrial and Applied Mathematics, in Portland, Oregon, US. “To us, it seems ‘wrong’”.
As an example, he pointed to a design for a lens that can be built from tiny cylinders of assorted sizes, which collectively focus the light passing through them.
First conceived of 70 years ago, the “human”-created versions of this lens use cylinders whose sizes are distributed symmetrically — an arrangement we tend to see as correct and elegant. But computerised efforts to improve that design wind up with a result that is anything but symmetric, with cylinders of varying sizes positioned according to no discernable pattern. “It’s not what you would expect,” Blankrot observes.
The initial impetus for such work, he says, comes from butterflies and other insects with vibrant colours. Such colours come not from pigments in their wings, but from microscopic structures that bend light.
But when physicists started trying to use the same approach they quickly discovered that the patterns in which structures are arranged can’t be perfect. Something has to break them up.
Furthermore such structures easily become extremely complex. “People can’t say off the top of their heads what will happen if I change this [element] or that one,” adds Blankrot.
So, his team handed it over to a computer, using an algorithm that told the computer if a change was in the “right” or “wrong” direction. It’s an approach he calls gradient-based optimisation – “which,” he explains, “basically means the computer algorithm knows which direction is best to go from the point you’re on.”
He uses the analogy of standing on a hill, not knowing how far you are from the top, but knowing which direction will lead you upward.
Blankrot isn’t worried that the results of such optimisations may sometimes appear incomprehensible. “It can be impossible to look at something that has 10,000 pieces and imagine how it will behave,” he says.
Nor is he worried that the computers doing the designs are about to take over the world.
“At the end of the day,” he says, “someone has to program this.” In addition, someone also has to review the results to make sure there aren’t errors in the code.
“There’s a lot of thinking that humans are good at,” he adds.
“Humans can give equations to the computer, and let the computer do the parts that are hard for us. We can stick to our strengths rather than doing the things the computer is good at.”
But not all scientists are equally prepared to join this brave new world.
Andrea Alu, an electrical engineer and physicist at City University of New York, notes that in many instances, it’s possible to calculate from theory the optimum performance that can be produced by any design, whether it’s created by humans or by “brute force” computerised methods.
If a concept is close enough to that optimum, he suggests, it doesn’t really matter if it comes from a computer algorithm or a human theoretician.
“Personally, I prefer a more physics-based approach,” he says. “I understand better when I see a solution that makes sense.”
But he agrees that computer-generated solutions are very popular, especially in optics, photonics, and acoustics. “It’s a very timely topic,” he says.