Deep Learning Processes for Rapid Quality Control from Wafer to Solar Cell

Successful quality control in the running production process requires fast and meaningful processing of extensive sensor data from different sources. For this, we develop multimodal deep learning models that enable quality assessment of the sample.

In numerous fields of application, processes are being developed that significantly improve the manufacturing, development and production of wafers and solar cells along the entire value chain. With our interdisciplinary expertise in artificial intelligence and photovoltaics, we implement task-specific AI models for this purpose. We have extensive measurement technology at our disposal, which is used to characterize the samples and "intelligently" train the models.

Two examples are presented below: 

Deep Learning for Contactless IV Measurement and Fast Measurement of Shingle Cells

© Fraunhofer ISE
Using an averaged IV measurement and an electroluminescence image of the host wafer (left), the quality of each individual shingle cell (right) can be precisely predicted.
Correlation graph between the contacted measured IV values and the predicted IV values determined using deep learning based on contactless measurements.
© Fraunhofer ISE
Correlation graph between the contact measured IV values and the predicted IV values determined by deep learning a set of non-contact measurements.

AI models are also used to develop alternative measurement methods. In the area of IV measurement, we have developed two approaches that have considerable potential to speed up and simplify the measurement:

  • AI methods have been developed to characterize half and shingle cells, which determine the quality of the individual shingle cells from measurements in the wafer composite (on the so-called “host wafer”) before the cutting process. This means that the time-consuming measurement of each individual shingle can be replaced. Compared to the pure IV measurement of the host cells, the AI model reduces the rejection of good partial shingles (see figure).

 

  • The most important parameter of a solar cell is the current-voltage characteristic. Until now, its measurement has required electrical contacting of the solar cell, which limits the throughput, subjects the cell to mechanical stress and requires regular maintenance of the contacting units. Our AI-based method enables contactless determination of the IV characteristic from a series of photoluminescence images and other optical measurements and thus represents a fast and stress-free alternative to contacted measurements, which is of particular interest for future high-throughput production (see figure).

The Digital Twin of the Solar Cell

© Fraunhofer ISE
Based on the production data, a data-driven digital twin of the solar cell is generated with a deep neural network. This enables efficient interaction with the experts as a "human-in-the-loop" and thus the definition of sorting criteria and research into the root causes of defects.

The process developed at Fraunhofer ISE for a data-driven digital twin addresses the challenge of evaluating and storing large amounts of data in production. Especially with multimodal data from different measuring systems, it is difficult for humans to recognize the usually complex patterns.

Our patented AI process for efficient data compression and data analysis offers an elegant solution to this problem: by correlating the high-dimensional measurement data with the final quality data of the cell, our system allows a data-driven fusion of the measurements. The result is a compact description of the sample that contains only relevant defect signatures and can be stored. This data-driven digital twin of the solar cell can be analyzed by the expert through efficient interaction with the "human-in-the-loop". With our AI support, root causes of defects can be efficiently identified and sorting criteria for production can be developed (see figure).

Our R&D services include:

  • development of AI models production
  • development of multivariate AI models for quality inspection from wafer to cell
  • transfer of AI models into the application: Defect detection with annotated data or based on our reference measurements using "smart labeling"
  • transfer and adaptation of the patented AI system for efficient data analysis, compression and storage of the measurements taken for production control purposes
  • development of operating concepts for the production line for quality inspection with just a few clicks and the "human-in-the-loop" using the digital twin
  • AI methods for fast, contactless IV measurement
  • AI methods for the inspection of shingle cells using measurements on host cells
  • transfer of AI models for quality inspection from photovoltaic production to other fields of application (such as fuel cell and battery production)

Research Project on the Topic Deep Learning Processes for Rapid Quality Control from Wafer to Solar Cell

 

NextTec

Contactless Performance Measurement for Solar Cells: Next Generation PV Production Technologies for Increased Throughput

Further Information on this Topic

Research Topic

Artificial Intelligence and Data Management

Business Area

Photovoltaics:​ Production Technology and Transfer