Using machine learning to find reliable and affordable solar cells
Researchers at the UC Davis College of Engineering are using machine learning to discover new materials for high-efficiency solar cells. They conduct complex experiments and apply various algorithms based on machine learning. As a result of the studies, they found it possible to predict the dynamic behavior of materials with very high accuracy without the need for a large number of tests.
The study was published in the ACS Energy Letters in April.
The object of the scientists' research is hybrid organic-inorganic perovskites (HOIPs). Solar cells based on hybrid organic-inorganic perovskites are a rapidly developing area of alternative energy. These molecules initiated the development of a new class of photovoltaic devices – perovskite solar cells. Their first prototypes were created in 2009.
Perovskites are comparable in efficiency to silicon for making solar cells, but they are lighter and cheaper to produce, which means they have the potential to be used in a wide variety of applications, including light-emitting devices.
However, there is an unresolved problem with perovskite-based devices. The issue is that they tend to break down faster than silicon when exposed to moisture, oxygen, light, heat, and stress.
The challenge for scientists is to find such perovskites that would combine high efficiency with resistance to environmental conditions. Using only trial and error methods, it is very difficult to quantify the behavior of perovskites under the influence of each stressor, since a multidimensional parameter space is involved.
The perovskite structure is generally described by the ABX3 formula, where:
A is a cation in the form of an organic (carbon-based) or inorganic group.
B is a cation in the form of lead or tin.
X is an anion, a halide based on chlorine, iodine, fluorine, or combinations thereof.
As you can see, the number of possible chemical combinations is huge in itself. Additionally, each of these combinations must be evaluated in multiple environmental conditions. These two requirements lead to a combinatorial explosion. We get a hyperparameter space that cannot be explored by conventional experimental methods.
As a first and key step towards solving these problems, researchers from the UC Davis College of Engineering, led by Marina Leite and graduate students Meghna Srivastava and Abigail Hering, decided to test whether machine learning algorithms could be effective in testing and predicting the effects of moisture on material degradation.
They built a system to measure the photoluminescence efficiency of five different perovskite films under repeated 6-hour cycles of relative humidity that simulate accelerated daytime and nighttime weather patterns based on typical northern California summer days. Using a high-throughput setup, they collected 50 photoluminescence spectra each hour and 7 200 spectra in a single experiment, that is enough for reliable analysis based on machine learning.
The researchers then applied three machine learning models to the datasets and generated predictions of environment-dependent photoluminescence responses and quantitatively compared their accuracy. They used linear regression (LR), echo state network (ESN), and seasonal auto-regressive integrated moving average with exogenous regressors (SARIMAX) algorithms and found values of the normalized root mean square error (NRMSE). Model predictions were compared with physical results measured in the laboratory. The linear regression model had NRMSE value of 54%, the echo state neural network had NRMSE of 47%, and SARIMAX performed the best with only 8% as NRMSE.
The high and consistent accuracy of SARIMAX, even when tracking long-term changes over a 50-hour window, demonstrates the ability of this algorithm to model complex non-linear data from various hybrid organic-inorganic perovskite compositions. Overall, accurate time series predictions illustrate the potential of data-driven approaches for perovskite stability studies and reveal the promise of automation – data science and machine learning as tools to further develop this new material.
The researchers note in their paper that generalizing their methods to multiple compositions can help reduce the time required to set up a composition, which is currently the main bottleneck in the design process of perovskites for light-absorbing and emitting devices.
In particular, the combination of SARIMAX with long short-term memory models (LSTMs) may allow prediction of perovskite chemistry beyond the training set, which will also lead to an accurate assessment of the stability of currently understudied compositions.
In the future, the scientists plan to expand their work by adding environmental stressors other than moisture (such as oxygen, temperature, light, and voltage). Combinations of many stressors can simulate operating conditions in various geographic locations, providing insight into the stability of HOIP solar cells without the need for lengthy experiments in each individual location.