Machine learning speeds solar energy research

  • August 10, 2021
  • Steve Rogerson

Machine learning can provide a shortcut to simulate interactions in materials for solar energy harvesting, according to researchers at the US Department of Energy’s Argonne National Laboratory.

Harnessing sunlight holds promise as a means to generate renewable energy cleanly for technologies such as solar fuel cells and water treatment systems. These require an understanding of what happens when materials and molecules absorb sunlight.

Computer simulations can help us better understand light-matter interactions. However, modelling materials featuring multiple types of structures, such as solid-water interfaces, is a complex task. But now, a research team at the Argonne National Laboratory has found a way to simplify these modelling tasks.

Using a data-driven approach based on machine learning, the team was able to simplify the solution of the quantum mechanical equations that describe how light is absorbed by a solid, liquid or molecule. Results of the research were recently published in Chemical Sciences.

“It is certainly not intuitive at first, but it turns out that machine-learning techniques can be used for purposes much different than recognising images or predicting consumer needs,” said Marco Govoni, co-author of the study and assistant scientist in Argonne’s materials science division.

The trick? Recognising that not all terms of the quantum mechanical equations need be computed in the same way. In fact, some terms could be calculated or learned from simpler quantities, speeding up the overall simulation.

“An important realisation of our work was to understand that we could reuse information obtained for a given solid or liquid without repeating calculations for similar systems,” said Sijia Dong, who was a postdoctoral fellow at Argonne when the research was conducted and is now assistant professor at Northeastern University. “In essence, we came up with a sort of recycling protocol to reduce the complexity of calculations required to simulate absorption of light by materials and molecules.”

These protocols can lead to big savings when it comes to simulations that may take many hours or even days on high-performance computing architectures.

In fact, the technique the team devised allowed simulations of absorption spectra of complex systems to run between ten and 200 times faster. These systems include solid-liquid interfaces such as those found between water and a photoelectrode, a material that can turn sunlight into electricity.

“Our study also gave insight into how to improve and modify the underlying theory used in the simulations,” said Giulia Galli, senior scientist in Argonne’s materials science division. “The impact of our machine-learning exercise turned out to be further reaching than expected; the data-driven approach we adopted indicated to us new ways to study light-matter interaction in even more realistic and more complex systems than the one we started out studying.”

The team is now looking at applying these shortcuts and recycling protocols to electronic structure problems not only related to light absorption, but also to light manipulation for quantum sensing applications.

Applications to silicon-water interfaces and water were funded by AMEWS, a research centre funded by the DoE. MICCoM funded the machine-learning algorithms as part of a computational materials science programme, supported by BES.