Scientists from UC Davis College of Engineering use machine learning to discover new materials for high -performance solar cells. They carry out complex experiments and use various algorithms based on machine learning. As a result of the research, they found that the dynamic behavior of materials with very high accuracy could be predicted without the need for a large number of tests.
The study was published in ACS energy lists in April.
The purpose of researching scientists is hybrid organic pervsky (HoIPS). Solar cells based on hybrid organic without organic pervskies are a rapidly developing area of alternative energy. These molecules initiated the development of a new class of solar devices – Perovskite solar cells. Their first prototypes were created in 2009.
Perovskites are comparable to the silo for the production of solar cells, but they are lighter and cheaper to produce, which means that they can be used in many different applications, including light emitting devices.
However, there is an unresolved problem with devices based on Perowski. The problem is that they tend to break down faster than silicon, when they are exposed to moisture, oxygen, light, warmth and stress.
The challenge for scientists is to find such pervsktles that would combine high performance with resistance to environmental conditions. By using only trial and error methods, it is very difficult to estimate the behavior of the Pervian under the influence of every stressor, because a multidimensional space of parameters is involved.
The Pervian structure is generally described by ABX3 Formula where:
AND It is a cation in the form of an organic group (based on coal) or inorganic.
B It is a cation in the form of lead or tin.
X He is an anion, a halogen based on chlorine, iodine, fluorine or its combinations.
As you can see, the number of possible chemical combinations is huge in itself. In addition, each of these combinations must be assessed in many environmental conditions. These two requirements lead to a combinatorial explosion. We get a hyperparametr space that cannot be examined by conventional experimental methods.
As the first and key step towards solving these problems, scientists from UC Davis College of Engineering, led by Marina Leite and PhD students of Meghn Srivastava and Abigail Hering, decided to check whether machine learning algorithms can be effective in testing and predicting the impact of moisture on the degradation of materials.
They built a system for measuring photoluminescence performance of five different films of Pervas under the repetitive 6-hour relative humidity cycles, which simulate accelerated weather patterns during the day and at night based on typical summer days in northern California. Using a high bandwidth configuration, they collected 50 photoluminescent spectra every hour and 7,200 spectra in one experiment, which is enough for reliable machine learning analysis.
Then scientists applied three machine learning models to data sets and generated forecasts of the environment -dependent photoluminescence answers and quantitatively compared their accuracy. They used linear regression (LR), Echo (ESN) and autoregressive integrated average movable medium with algorithms of exogenous regressors (SARIMAX) and the values of the normalized average square error (NRMSE) were found. Model forecasts were compared with physical results measured in the laboratory. The linear regression model had a NRMSE value of 54%, the echo neural network had 47%NRMSE, and Sarimax worked best with only 8%as NRMSE.
The high and consistent accuracy of SARIMAX, even when tracking long-term changes in a 50-hour window, shows the ability of this algorithm to model complex non-linear data from various hybrid organic organic compositions of pervskt. In general, accurate forecasts of time ranks illustrate the potential of approaches based on data for stability research and reveal the promise of automation-nauki data and machine learning as tools for further development of this new material.
Scientists note in their article that generalizing their methods for many compositions can help reduce the time needed to configure the composition, which is currently the main bottleneck in the process of designing Pervian to devices for playing light and emitting.
In particular, the SARIMAX combination with long short -term memory models (LSTMS) can allow for anticipation of Perovian chemistry outside the training set, which will also lead to a thorough assessment of the stability of currently underaided compositions.
In the future, scientists plan to expand their work by adding environmental stressors other than moisture (such as oxygen, temperature, light and tension). Combinations of many stressors can simulate working conditions in various geographical locations, ensuring insight into the stability of HOIP solar cells without the need for long experiments anywhere.
