Improving Industrial Solar Cell Fabrication using Neural Networks and Genetic Algorithms

Improving Industrial Solar Cell Fabrication using Neural Networks and Genetic Algorithms

Eijkens, Casper (TU Delft Electrical Engineering, Mathematics and Computer Science)

Isabella, O. (mentor)
Hameiri, Ziv (mentor)
Buratti, Yoann (mentor)

Delft University of Technology

2019-09-23

The present thesis approaches the improvement of the performance metrics of industrially fabricated solar cells. It aims to develop an optimization method that uses machine learning to identify an improved configuration of solar cell production lines. Conventional methods for process improvement rely on modelling experimental data to improve a process of interest. The number of experiments needed for modelling grows exponentially with the number of parameters taken into account. Since experiments are costly and time-consuming, the number of parameters that can be considered in this existing method is limited. This existing method thus allows modelling of only a few processes, and cannot include all inter-dependencies between the processes in the production line. In this thesis, an alternative method is proposed for the optimization of solar cell fabrication that uses machine learning and genetic algorithms. Instead of collecting experimental data, this method uses data collected from sensors in the production line. Machine learning is employed to establish a model of the production using the natural variation in this sensor data. This allows us to create more complete and accurate models of the production line. Genetic algorithms are then combined with the machine learning model to identify an improved configuration of process parameters. As part of this study, a virtual solar cell production line was developed to test and compare different machine learning algorithms. Based on experiments with this virtual production line, two novel approaches are proposed that can optimize production lines using a combination of artificial neural networks and genetic algorithms. It was demonstrated that the first approach improved the photovoltaic cell efficiency of cells from a simulated production line from 17.9$\pm0.3\%$ to 19.2$\pm0.2\%$. This significant improvement demonstrates the power of applying machine learning to solar cell production. The results achieved in this thesis encourage further development of the proposed method and application to real-life production lines.

Photovoltaics
Machine Learning
Design of Experiment
Process Improvement
Simulation

http://resolver.tudelft.nl/uuid:3787ea40-dae4-4841-a101-05a5c0acd266

Student theses

master thesis

© 2019 Casper Eijkens