‘Design Mining’ Reimagines Engineering with Evolution and Neural Networking

A pair of researchers from the University of West England wants to make better wind turbines through a process called design mining. Never heard of it? This is a scheme in which mathematics and resource-intensive fluid dynamics simulations are tossed in favor of something like biological evolution, with help from 3D printing. It's a way of letting technology design itself.

Imagine that at the beginning of the design process is just some best guess, what amounts to a design seed. This turbine design-seed (my term) is tested in real life using some fans and what works stays and what doesn't is scrapped. The process repeats, with each iteration keeping the good and scrapping the bad. The end result might just be the same thing you'd get through the usual process, but it also might be something totally weird and unexpected. Evolution is good like that, and maybe it's a way of doing things that can be harnessed by engineers.

So it's not unlike natural selection. Mutation occurs and if it works it has a better chance of persisting through generations. For something like a wind turbine farm, selection via design-mutation might even be a more suitable process, just given the complexities and unpredictabilities that arise as turbines interact within a windmill farm. It's a situation that keeps windmills more widely dispersed than may actually be needed and, thus, farms less efficient.

The West England duo, Richard Preen and Larry Bull, explain it as such. "Candidate prototypes are evaluated under fan generated wind conditions after being physically instantiated by a 3D printer," they write. "That is, unlike other approaches, no mathematical formulations are used and no model assumptions are made. Initially, artificial evolution was used to explore the design space of a single isolated [vertical-axis turbine] and subsequently a cooperative coevolutionary genetic algorithm was applied to explore the design space of an array of 2 closely positioned [turbines]."

It's a technique that draws on computer science as much as biology; or, rather, computer science can wind up looking like biology and vise versa more often than one might think. Preen and Bull are describing an iterative method for design, in which a problem is solved basically by chipping away at it in successive loops until some condition is met.

The pair point out that one of the big advantages of 3D printing (or rapid-prototyping) is that it's possible to fabricate something directly into an environment. Here, that environment is a miniature wind farm. They note also that evolutionary design isn't in itself a brand new idea; in 2011, for example, researchers came up with a new design for NASA's ST5 spacecraft using evolutionary methods. But doing this within a physical space puts a new twist on the concept.

The evolutionary selection process is further boosted by neural networking/machine learning techniques, in which a single "meta" model or prototype is refined through time as more and more training data is collected. Again, a bit of biology and a bit more computer science. Initial results are promising: evolutionarily-generated designs ("fittest" designs) have handily beat out conventional human-designs, at least within experimental conditions.

"The use of 3D printing to physically instantiate candidate designs completely avoids the use of 3D computer simulations, with their associated processing costs and modelling assumptions," Preen and Bull write.

"In this case, 3D [computational fluid dynamics] analysis was avoided, but the approach is equally applicable to other real-world optimisation problems, for example, those requiring computational structural dynamics or computational electromagnetics simulations," the continue. "We anticipate that in the future such 'design mining' approaches will will yield unusual yet highly efficient designs that exploit characteristics of the environment and/or materials that are difficult to capture formally or in simulation."