General deep learning framework autonomously selects suitable materials from self-built library and optimizes structural parameters in designs.
Designing materials and structures with a specific emissivity — known as wavelength-selective thermal emitters (WS-TEs) — can be a difficult engineering challenge for applications including energy harvesting, radiative cooling and thermal camouflage.
Depending on the application, the tendency has been either to fix materials and design structures or fix structures and select suitable materials. The result is suboptimal designs for WS-TEs.
However, a recent paper from a team of engineers and researchers at Huazhong University of Science and Technology has proposed a new general deep learning framework for optimizing WS-TE design across emissivity engineering applications. The team used this framework, based on the deep Q-learning network (DQN) algorithm, to design multilayer WS-TEs for several applications, including thermal camouflage.
Different applications generally require WS-TEs with different emissivity. For example, thermal camouflage needs low emissivity in the long infrared range to prevent detection while radiative cooling needs high emissivity so that the thermal power can be released at a lower temperature of the device. Since the emissivity of WE-TEs is contingent on both their materials and their structure, designers need to balance these factors carefully, which can be difficult for a human or even a team of humans.
The DQN algorithm solves this problem by autonomously selecting materials from a self-built library while simultaneously optimizing the designs’ structural parameters based on the target emissivity spectra of each application. The researchers reported in Light: Science & Applications that the three WE-TEs they fabricated based on the algorithm’s designs showed excellent performance and matched the target emissivity in each case.
“The input parameters of the DQN framework are highly flexible in materials, structures, dimensions, and the target functions, offering a general solution to other nonlinear optimization problems beyond emissivity engineering,” write the researchers.
This is a perfect example of how artificial intelligence (AI) can augment engineers in their day-to-day work. As the researchers point out, even with just eight available materials for WE-TEs, there are 1.75 x 1010 potential structures for designs. The sheer volume of the optimization space makes manual design impractical, at best.
Rather than spending their time trying to balance materials and structures, engineers could use the DQN framework to find solutions automatically.
But the really intriguing question is how far this sort of frame can be applied. Nonlinear optimization problems are an engineer’s bread and butter, so what happens when they can be solved with an algorithm?
Even in that radical possibility, there’s still an obvious place for engineers, namely on either side of the AI in the workflow. Engineers would still need to provide the DQN algorithm (or its successor) with emissivity targets and other considerations. And they would certainly still need to evaluate its designs. In the end, it’s not so different from many of the other software tools engineers use every day.
