Junichiro Shiomi and co-workers from The University of Tokyo have applied materials informatics and machine learning (ML) to determine the optimum nanostructure that would minimize lattice thermal conductivity by phonon engineering. Such minimizing of thermal conductivity would be useful in thermal barriers and thermoelectrics. The key to their success is the ability to tune coherent phonon heat conduction.

Thermal phonons, when scattered by nanostructures, are usually considered as incoherent particles due to their relatively small wavelengths. The nanostructures typically diffusely scatter thermal phonons, yielding a reduction in the lattice thermal conductivity. Coherent phonons on the other hand have revealed much lower lattice thermal conductivities, since they are considered to have wave-like properties (wave-particle duality), allowing for phonon wave interference. This allows for manipulation of wave transport and thus tuning of thermal conductivity.

“While the phonon particle picture has been leveraged for nanoengineering in the past few decades, phonon engineering based on its wave picture has emerged to attract increasing attention,” says Zhiting Tian, associate professor at Cornell University, who was not involved in this study. “This work combined computational and experimental work to demonstrate the power of manipulating coherent phonon transport in aperiodic superlattices,” she says.

Superlattices represent a suitable material for exploring phonon coherence due to the presence of interfaces that are smooth. “Interface smoothness is quite important to tune coherent phonon behavior,” says principal investigator Junichiro Shiomi. “Reflection and transmission need to be specular, not diffusive, and for that, the surface roughness needs to be smaller than the wavelength of phonons,” he says. Shiomi tells MRS Bulletin that an important characteristic of phonons is that their average wavelength is short—typically only a few nanometers at room temperature, so the interface needs to be smooth. 

In order to investigate the phonon coherence effect, Shiomi and colleagues built GaAs/AlAs superlattices that were “synthesized with nanometer layers and coherently bonded interfaces,” as they reported in a recent issue of Physical Review X. Through molecular beam epitaxy, the researchers grew a superlattice film with a thickness of 16 unit layers on a GaAs substrate. Each of the ~4.5-nm thick unit layers is either GaAs (designated by 0) or AlAs (1), yielding 2¹⁶ or 65,536 possible structures.

Machine-learning (ML)-based materials informatics is emerging as a new paradigm in materials science for developing new materials. In order to optimize the structure of the superlattices in this work, the researchers utilized ML materials informatics from which they derived an optimal aperiodic structure. This was identified from a total of 65,536 candidate structures to minimize thermal conductivity by manipulating coherent phonon transport. ML optimization took about 3-4 days, which can be further shortened by parallelization. The researchers verified the desired thermal conductivity by measuring it with the time-domain thermoreflectance technique in the temperature range of 77-300 K.

In a comparison with a periodic superlattice consisting of alternating GaAs and AlAs unit layers, the optimized aperiodic superlattice (of 1001010101101101) expressed a smaller thermal conductivity. Upon further examination with an atomic-resolution transmission electron microscope, the researchers could view the near-perfect interface with monolayer roughness.

According to the researchers, the consistent thermal resistance across the 77-300 K range means that the wave interference of coherent phonons can be controlled. “Since a periodic SL [superlattice] has more GaAs/AlAs interfaces,” write the researchers, “the superiority of an aperiodic SL suggests successful engineering of coherent phonon heat conduction.”

“I think it is an elegant way to incorporate machine learning to help optimize the aperiodic superlattices, which is otherwise impossible,” says Tian. “Although such optimization has been applied to nanostructures for desired properties before, the beauty of this work is to validate the prediction with experiments and further analyze the underlying physics—localization and local patterns,” she says.

Shiomi and co-workers have explored similar ML-optimization of nanostructures for holey graphene for thermoelectrics and disordered MgAl2O4 for magnetic tunnel junctions, which were both theoretical studies; and over a year ago he published a similar study that was demonstrated experimentally on multi-layered photonic metamaterials for wavelength-selective thermal emission. Shiomi tells MRS Bulletin that it is easier to tune the coherence for photons because the wavelength is much larger (in the range of micrometers).

Tian says, “This [current study] gives us more confidence in machine-learning-based materials informatics.”

Read the article in Physical Review X.