A dynamic calibration method for injection-dependent charge carrier lifetime measurements
A dynamic calibration method for injection-dependent charge carrier lifetime measurements
A dynamic calibration method for injection-dependent charge carrier lifetime measurements
Machine learning applications for photovoltaics
Parameter exctraction from simulated curves
We develop machine learning (ML) techniques to extract key performance parameters from simulated curves.
Traditional methods for analysing internal quantum efficiency (IQE) measurements are slow and manual. They often require complex fitting, especially for non-silicon technologies like gallium arsenide. Our deep learning models automate this process. They accurately predict multiple electrical and optical parameters from IQE data, even when the data is noisy.
We also use ML to identify and extract bulk defect parameters in high-efficiency silicon solar cells. Our models are trained on large simulated datasets. They predict defect energy levels and capture cross-sections with high precision. They can also distinguish between one- and two-level defects with over 90% accuracy.
Automatic quantitative analysis of internal quantum efficiency measurements of GaAs solar cells using deep learning. A representative example of multiple solutions providing a similar fit; a) the true IQE, convolutional network (CNN) prediction, and manual fits. b) The RMSE heatmap, with the color bar set to a maximum of 1.5 × 10−3. The red star identifies the true values while the blue open diamond denotes the CNN model's prediction.
Luminescence image pre-processing
We also apply ML techniques to enhance and process luminescence images for solar cell analysis.
Luminescence imaging is a fast, non-destructive method to detect spatial defects and assess electrical properties in photovoltaic devices. However, achieving high-resolution and low-noise images often requires expensive equipment or long exposure times. Our deep learning methods improve image quality by enhancing spatial resolution and reducing noise—without the need for costly hardware upgrades. This makes it easier to detect small defects and extract accurate performance data. Our models improve image clarity by over 30% in peak signal-to-noise ratio and 39% in structural similarity.
We also address perspective distortion in images taken by drones. Our models estimate the camera angles, including yaw, pitch, and roll. They then adjust the images to correct the distortion.
Deep learning model to denoise luminescence images of silicon solar cells. Denoising of the EL image: a) a representative original EL image, b) a noisy image, c) a BM3D-reconstructed image, and d) a U-net-reconstructed image.
Cell defect classification
We also design ML models to identify and classify defects in electroluminescence (EL) images of solar cells.
EL imaging is widely used in solar cell production to detect performance losses. It reveals defects caused by recombination and high series resistance (Rs). However, separating these two types of defects is challenging.
Our deep learning model extracts qualitative photoluminescence (PL) and Rs images from a single EL image. It is trained on simulated datasets and tested on unseen simulations and experimental measurements. Preliminary results demonstrate that our model can accurately distinguish whether defects are related to recombination or series Rs losses.
Extraction of qualitative PL and Rs images from a single EL measurement. Representative images. (a-c) show the ground truth EL, PL, and Rs images, respectively. (d-e) present the model's predictions for PL and Rs, respectively, using the ground truth EL image as an input.
Efficiency loss analysis
Beyond defect classification, we develop advanced techniques to analyse the impact of defects on solar cell efficiency.
Identifying efficiency shortfalls in production lines is essential for improving manufacturing processes. We propose a method that uses generative adversarial networks (GANs) to automatically reconstruct defect-free luminescence images. By comparing the reconstructed image with the original, we can estimate the efficiency loss caused by each defect. This approach has been validated on intentionally damaged cells, demonstrating accurate efficiency loss prediction.
Our method not only detects and localises defects but also quantifies their individual impact on efficiency. Applied at scale, this technique helps manufacturers quickly identify defect patterns that lead to the greatest efficiency losses.
​
Automated efficiency loss analysis by luminescence image reconstruction using generative adversarial networks. A defect localisation algorithm (2) and EL reconstruction algorithm (3), both based on GANs, identify defects in an EL image and generate a defect-free image. Both original and reconstructed cell efficiencies can be assessed using an efficiency-prediction algorithm (1), thus, assessing the efficiency shortfall due to the presence of the defect.
End-of-life module assessment
We also apply ML to assess the condition of photovoltaic (PV) modules in the field.
End-of-life (EoL) management is a growing challenge for PV systems. Luminescence imaging is often used to assess module quality. However, current methods are manual and slow. They rely on experts to inspect each image, which limits their use at scale.
We have developed an automated framework for PV module assessment. It processes luminescence images and identifies defective cells within each module. The system includes image pre-processing, defect detection, and classification. It also estimates power loss and remaining power generation potential.
The results help operators make informed decisions about EoL actions. These include repair, reuse, or recycling. The process is fully automated, scalable, and cost-effective. It supports large-scale PV installations and improves the efficiency of EoL management.
​
Automated photovoltaic module quality assessment. Example from luminescence image after pre-processing and segmentation (analysis performed on individual cell images); identified defects are labelled by class (none, damaged or defective, dead or degraded, etc) and severity (coloured by increasing risk: yellow, orange, and red).