Skip to main content

Modeling, machine learning optimize nanostructure of thermoelectric graphene nanoribbons

By Melissae Fellet August 23, 2018
The optimal structure of a thermoelectric graphene nanoribbon contains an aperiodic pattern of holes. (a) The optimal structure has a higher figure of merit (ZT), power factor (P), and thermal conductance (Rth)—properties that reflect thermoelectric ability—than a pristine graphene nanoribbon or one with a periodic arrangement of holes. (b) Periodic (top) and optimal aperiodic (bottom) structures. Credit: Science Advances

A combination of modeling and machine learning has been able to identify the optimal nanostructure for graphene nanoribbons for a thermoelectric to convert heat to electricity. This optimization strategy could be applied to other materials where a desired physical property requires balancing conflicting characteristics.

Thermoelectric materials could be used to power sensors and transmitters. The best thermoelectrics are nanostructured materials that have low thermal conductivity, so they do not wick their energy source away. They also have high electrical conductivity for electricity produced to flow easily. However, materials that conduct electricity well often conduct heat well too.

Computer modeling can identify the best nanostructures for optimal thermal or electrical conductivity. But it is harder for a model to identify one structure that satisfies multiple requirements that inherently require a functional trade-off, as in the case of thermoelectric behavior. This is where machine learning comes into play.

Junichiro Shiomi, at the University of Tokyo, and his colleagues wondered if they could use machine learning to balance the conflicting demands inherent in thermoelectric material design. For their model system, the researchers chose a graphene nanoribbon containing 16 sections that may or may not contain a 2.8 Å hole. Under those conditions, 32,896 structures with varying numbers and arrangements of holes were possible.

Since there were too many structures to model each one efficiently, the researchers first chose a random sample from that population. They used Green’s functions to calculate the thermal and electrical conductance of the randomly selected structures. The researchers then used that conductance data to train a machine learning algorithm.

Machine learning can find mathematical relationships between data that are seemingly unrelated. It has been used to power email spam filters, virtual personal assistants, and automated video surveillance. Machine learning algorithms train with a set of known information and then build a model that connects the various inputs. The result is that the program learns its task without explicit programming.

To efficiently build the training set for the thermoelectric nanoribbon structure optimization, the researchers selected the best nanoribbon structures predicted by machine learning. They then calculated the thermal and electrical conductance of these structures and added that data to the machine learning optimization.

The researchers repeated this alternating modeling and machine learning process about 100 times. Eventually, a graphene nanoribbon with a seemingly random arrangement of holes emerged as the optimal thermoelectric material. “I would have never identified this structure intuitively,” Shiomi says. The optimal material had a figure of merit, ZT, 11 times greater than that of pristine graphene nanoribbons without holes.

Prasanna Balachandran, at the University of Virginia, says the computational materials science community as a whole is recognizing that it is important to do calculations with intelligent sampling rather than brute force, calculating a property for all possible structures. The method presented in this paper is a way to tackle a large parameter space, he says. The next step would be to validate these simulations with experiments, he adds.

Read the abstract in Science Advances.