top of page
Machine learning applications for photovoltaics

Extraction of defect parameters from TIDLS measurements using machine learning

The performance of high-efficiency silicon solar cells is limited by the presence of bulk defects. Identification of these defects has the potential to improve cell performance and reliability. The impact of bulk defects on the minority carrier lifetime is commonly measured using temperature and injection-dependent lifetime spectroscopy (TIDLS). The Shockley-Read-Hall equation is then often used to fit the measurements and extract the defect’s parameters, such as its energy level and capture cross-section ratio.
In this project, we are developing an alternative extraction approach by using machine learning models trained on more than a million simulated lifetime curves. With no a-priori knowledge of the physical equation, the machine learning models achieve coefficients of determinations between the true and predicted values of the defect parameters above 99%. The successful application of machine learning in the context of bulk defect parameter extraction paves the way to more complex data-driven physical models which have the potential to overcome the limitation of traditional approaches. They achieve this by incorporating the temperature dependence of the capture cross-section and the presence of multiple single-level defects. The developed method is not limited to silicon and will be applied to other materials such as perovskite and CIGS.


Defect parameter regression using random forests for both (top) defect energy level (Et) and (bottom) capture cross section ratio (k).

Sorting solar cells using luminescence imaging and machine learning

End-of-line binning of solar cells ensures optimal power output of photovoltaic modules, as well as identification of misprocessed cells. Currently, binning is performed using current-voltage measurements. In this project, we are developing a deep learning framework to detect defective cells, predict cell efficiencies and bin the cells directly from electroluminescence imaging using a custom-made convolutional neural network. Already at this early stage, the network identifies defective cells with 93.7% accuracy and predicts cell efficiency within a mean error of 0.15% absolute efficiency. This framework can be also applied to half-cut cells, providing a solution for post-cutting binning of this new structure.
This project paves the way to deep learning applications in solar cell production lines and unlocks the potential of luminescence imaging as the ultimate end of line process monitoring and quality control tool.

Sorting ML.png

True vs predicted efficiency using our developed deep learning framework.

Photovoltaic module fault prediction using luminescence imaging and machine learning

Using innovative combinations of luminescence-based imaging and machine-learning, this project aims to extend the durability and reliability of photovoltaic modules by early-detection of degradation processes. Using luminescence images of modules and machine-learning algorithms, automated methods to identify module degradation mechanisms will be developed. This will significantly improve the durability and reliability of photovoltaic systems. It is estimated that ~20% of modules installed in Australia will under-perform, causing substantial risk for the financing of photovoltaic systems and ultimately hurting Australian consumers. Extending the capabilities of luminescence images for early identification of faulty modules will reduce that risk and lower the cost of photovoltaic power-plants.

Energy Efficiency Consultation

Solar Photovoltaic Array

Optimisation of solar cell production using machine learning

To keep improving the performance of photovoltaic solar cells, manufacturing lines need to be continuously optimised to produce the highest possible cell efficiencies. Usually, a process-wide optimisation is performed, as simultaneously optimising all the processes is costly and time consuming.
In this project, we are developing a machine-learning-based method using the natural variation of a production line, to investigate the relationship between input process parameters and output cell efficiency. A genetic algorithm is then employed to identify new process parameters that maximise the cell efficiency. By implementing those new process parameters, a new cell efficiency distribution is created, and the same approach can be repeated.
We have already tested the proposed method using a simulated production line of mono-crystalline aluminium-back surface field solar cells. Using a neural network, an accurate model is built to predict cell efficiencies from input process parameters with errors less than 0.03% absolute efficiency. The method is then used to increase the mean cell efficiency from 18.07% to 19.45% in just five optimisation iterations. Provided strong process monitoring and accurate wafer tracking, the proposed method is applicable, as is, to production-type data sets, enabling the photovoltaic industry to build smart factories and join the fourth industrial revolution.


Successive iteration of the proposed approach, showing the predicted  and true efficiency of the average of each generation followed by the efficiency distribution generated by the new recipe found. Each dotted line reports the maximum efficiency of the distribution, generated from the previous step.

bottom of page