Photovoltaic solar – what most people know as solar panels – generated a record 821 terawatt-hours in 2020, a 23 percent increase over 2019 and an impressive 3.1 percent of global electricity generation. But with climate change rapidly accelerating, most scientists agree that the energy transition needs to be happening faster – which means producing more, cheaper and more efficient renewable energy. Now, a team of researchers led by Coen de Graaf of the University of Rovira i Virgil (URV) in Spain is investigating new organic materials for photovoltaic panels that could provide benefits like lower costs and lighter weights. To assist in that materials research, they’ll also be employing two of the world’s most powerful supercomputers: Oak Ridge National Laboratory’s IBM-built Summit system (the top-ranked supercomputer in the U.S.) and Jülich Supercomputer Center’s Atos-built Juwels system (the top-ranked supercomputer in Europe).

Investigating a step change in efficiency

De Graaf and his team are aiming to produce a better understanding of how certain materials engage with the processes at the heart of photovoltaic power production. “Our project is rooted in basic research. It does not directly aim at the design of immediately applicable materials but rather at understanding the fundamental processes that are known to boost the efficiency of photovoltaic cells based on organic materials,” explained de Graaf, a professor in experimental sciences and mathematics at URV, in an interview with HPCwire. “The materials we focus on are medium-sized organic molecules such as tetracene and pentacene. These molecules are capable of absorbing sunlight and generating not just one electron-hole pair as is the common mechanism, but rather generate two electron-hole pairs.”

This process, de Graaf continued, is called singlet fission and could theoretically (though, he cautioned, not realistically) double the efficiency of a solar cell through the use of such organic materials. “We are scanning all these new materials on the properties that determine the potential efficiency for singlet fission,” he said. “This is done by simulating the electronic structure of these molecules.”

Simulating those structures, though, is easier said than done, with de Graaf explaining that many of the methods used to simulate electronic structures of molecules “can only partially address the properties relevant for singlet fission.” Instead, the team is turning to a method called non-orthogonal configuration interaction (NOCI) that was developed in the ‘80s, but which fell into disuse due to high computational costs – until recently, when it was picked back up through an open-source code called GronOR.

But even though computers have advanced quite a bit in the last forty years, NOCI is still “computationally very expensive,” de Graaf said. “However, the method is extremely well-suited for parallelization and a large part of the process can relatively easily be done on GPUs.”

Computational firepower

So, with those needs in mind, the team obtained time allocations on two gargantuan systems. First, an 1,080,000-node-hour allocation on the 148.6-Linpack petaflops Summit supercomputer via the U.S. Department of Energy’s INCITE (Innovative and Novel Computational Impact on Theory and Experiment) program; second, a 243,550-node-hour allocation on the 44.1-Linpack petaflops Juwels booster module (pictured in the header) via PRACE (the Partnership for Advanced Computing in Europe). Both systems pack a sizable number of GPUs: 27,648 Nvidia Volta GPUs in Summit, 3,744 Nvidia A100 GPUs in the Juwels booster module.

“[The] two architectures that we use are just what we need for making our calculations possible,” de Graaf said. “GronOR scales linearly up to the machine limit on both machines and can reach a speed-up of more than 20 when the GPUs are used. … One of the largest calculations that we have done so far was a run of 4,600 Summit nodes, each with 6 GPUs, and each GPU holding three ranks, leading to more than 80,000 processes running simultaneously. The Juwels computer has four GPUs per node, but being of a newer generation, our calculations on both machines are approximately equally fast.”

Asked about the timeline for the project, de Graaf said that the INCITE allocation on Summit spans two years and begins in January. The PRACE allocation on the Juwels booster, meanwhile, is a one-year project that began in October.

“Now with this massively parallel and GPU-accelerated program, NOCI calculations can be performed for molecules with more than 100 atoms, which opens the possibility to screen many molecules with potentially interesting singlet fission properties,” de Graaf said. “This will not only let us propose new candidates for the layer in the organic solar cells where sunlight is absorbed and charge carriers generated, but also (maybe more importantly) understand how chemical modifications affect the singlet fission properties, and hence, lead to design rules for new materials.”