Macrocosm: Machines Are Learning To Make The Most Efficient Solar Cells
thianchai sitthikongsak
Many experts see alternative energy sources as the only way to proceed into the future. Non-renewable sources like coal, oil, and natural gas release billions of tons of carbon dioxide and other greenhouse gases every year. 2019 marked a year of the most global carbon emissions in history, as did the two years before it. Increases in greenhouse gas emissions contribute to higher temperatures, rising sea levels, and heavier periods of natural disasters, such as hurricanes and wildfires. It's clear that humanity must curtail these sources of emissions; however, in order to keep up with a rapidly rising demand for energy, other sources must be found.
One of those alternative energy sources sits 91 million miles from the earth, keeping us in a constant orbit. Solar power is energy from the sun in the form of electrical or thermal energy. Humans have captured solar energy for thousands of years. The first documented use of solar power was by the ancient Egyptian ruler Amenhotep III, who supposedly possessed statues that operated when the air in their base pedestals expanded after exposure to sunlight. Though this was merely for aesthetic effect, records of ancient Egyptian and Greek homes show that many older civilizations used passive solar energy to circulate air and distribute heat through their homes. Anasazi Native Americans built their homes on the south sides of cliffs to make use of the sun during the winter while avoiding its harsh rays during the summer.
Modern solar energy uses mostly pertain to the utilization of small particles of sunlight to knock electrons from atoms, which generates a flow of electricity. This process involves the use of solar panels, large surfaces of photovoltaic cells that can absorb the energy of light and transfer that energy into electricity. These photovoltaic cells are often made of silicon, a crystalline structure that is thick, largely inflexible, and must be incorporated in very rigid ways. The large swatches of uniform sunlight-facing panels often seen on roofs or in fields are due to this strict requirement. Silicon solar panels simply do not work in any other way. Since they were the first on the market for commercial solar panel use, this has been accepted into the mainstream as the "only way" to have solar energy.
However, there is another way to gather solar power -- one that has recently gotten a boon from machine-learning.
Organic Solar Cells
Contrary to commonly held belief, organic, in a chemical sense, neither means 'naturally-occurring' nor 'healthy to eat.' Organic refers to the presence of the element carbon, which is found anywhere from Tupperware to textiles to toothbrushes. Carbon is beneficial due to its ability to make multiple bonds, which is likely why all life on earth evolved to utilize it for most biological functions. Carbon is flexible, light, and adaptable, which differentiates it from silicon, it's neighbor on the periodic table.
These properties of carbon also make solar cells made from them different from the typical silicon solar cells. Organic photovoltaics (also known as OPVs) are made with compounds that are typically dissolved in ink and printed onto thin plastics, allowing OPVs to be incorporated into more places and structures than silicon photovoltaics. New, unobtrusive solar power walls and windows are made from OPVs. This flexibility would allow those who don't have large swathes of roof spaces or fields to use solar panels to provide electricity for homes, businesses, and community buildings. For urban environments that boast little available space, flexible OPVs could be a groundbreaking new solution to the current energy crisis.
OPVs work through high-performance photoactive materials that can capture light particles called photons and knock electrons loose from a semiconducting material. This process allows an electrical current to flow. This requires one material to be a donor of electrons and one material to be an acceptor of that electron, like two edges of land that support a bridge's ropes. Only certain materials can act as donors and acceptors, and many have not been discovered yet. In the last decade, the discovery of new donor and acceptor materials has accelerated the design and synthesis of tens of thousands of photovoltaic materials. This discovery is vital to developing cheaper, more flexible, and more efficient organic solar panels.
A current barrier to implementing OPVs is their efficiency. According to the United States Department of Energy, organic solar cells exhibit efficiencies of around 10 percent, whereas crystalline silicon solar panels easily double that. An inefficient solar panel is not an endeavor that private and public sector entities want to invest in. Organic solar panels are not commercially available for that very reason. If OPVs want a place in future commercial sales, they'll need to grow far more efficiently.
And that's where AI comes into the equation.
Machine’s Role in Harnessing the Sun
A recent study by scientists at Wuhan University and the Beijing National Laboratory for Molecular Sciences tested the ability of a machine-learning model to collect data regarding donor/acceptor pairs, within parameters that enabled them to work together. It then predicted which materials, out of a given library, would work together as donors and acceptors.
The study highlighted two predictive approaches that the machine used to pick these donor/acceptor pairs: random forest (RF) and boosted regression trees (BRT). Random forest classification consists of a large number of individual decision "trees" that operate together. Each tree will make a prediction and the forecast with the most amount of trees "voting" for it becomes the model's prediction. This predictive approach is beneficial for minimizing error, as it does not rely on a single tree to make a prediction. An uncorrelated model makes every prediction. The trees don't interfere with each other and produce a cumulative forecast that is more accurate than any individual prediction.
Boosted regression trees make use of these trees of decision-making and take into account the previous trees. Some are "boosted" or weighed differently in subsequent calculations. If previous trees poorly modeled the data, it has a higher probability of being selected in a new tree. This technique takes into account error, the complexity of the tree, and the machine's learning rate as it adapts and assigns weight to different outcomes.
Using both of these techniques, the model screened over 32 million pairs of donors and acceptors. Six couples were selected and synthesized to compare their predicted efficiency to the experimental efficiency. The experimental efficiency showed a high agreement with the anticipated results, signaling to the scientists that the machines could pick out six highly efficient pairs from their extensive database.
These pairs of donors and acceptors could be among those used in the mainstream sale of organic solar cells. Or perhaps the next batch of statistical analysis by a machine will bring us the undiscovered gem of highly-efficient, highly-flexible, and inexpensive material that could produce the next generation of solar energy. This study proves one thing: machines can learn how to best capture sunlight from the air and transform it into power.