Scientists in Korea have developed a new method to use machine learning models in the production of “smart” solar cell production. They used data collected from equipment that looks strongly like actual industrial production tools.
Researchers from Korea University have developed a machine learning model for predicting leaf resistance in doping processes of phosphoroxychloride (POCL3) in the production of solar cells.
“Our study is intended to present a methodology for integrating machine learning in industrial processes, with the aim of speeding up the progress of Industry 4.0 and to clear the way to industry 5.0,” the main author of the study, Seungtae lee, said PV -Magazine.
“We used data collected from equipment that resembles actual industrial production tools,” he said. “With the help of this data we have developed a Machine Learning model, not only to predict the leaf resistance based on process conditions, but also to optimize those conditions by Bayesian optimization to meet specific goals of the leaf resistance.”
In the production of solar cells, POCL3 is used as a liquid dopant leader to create n-type layers in the thermal diffusion process.
For their modeling, the scientists considered different oven process conditions and leaf resistance values. They collected 3,420 experimental data points, in which 10 process variables were used as input parameters: temperature for deposition; Pre-deposits; Drive-in circumstances; drive-in temperature; Drive-in time; process gas parameters; POCL3 -current speed, O2 -current speed and process pressure; Wafer boot position; Wafer final number; And waffle position.
The research group used the Shapley Additive Expanations (SHAP) method, a game theory approach to explain the output of each machine learning model to analyze the impact of each function on the forecast of leaf resistance. “Shap is an interpretation technique based on Shapley values from game theory,” it emphasized. “It offers an extensive quantitative analysis that includes the importance of functions, the influence of each characteristic on model predictions and the specific contributions of individual characteristics to any prediction at the data point level.”
The academics also used Bayesian optimization, which is often used to solve complex optimization problems by approaching an unknown objective function and efficiently identify the minimum or maximum values, to identify the optimum process conditions by using the trained machine learning model. In particular, they tried to identify the circumstances that yield a sheet resistance, almost 150 Ω/SQ under “realistic” production conditions of solar cells.
The team carried out 100 tests in the initial random sampling phase and 100 tests in the Bayesian optimization phase.
The proposed approach appeared to achieve a more efficient and rapid optimization of process conditions compared to conventional and expensive test-and-error methods used in the PV industry.
“We have found that the learned representations and predictions of the model are consistent with established physical and theoretical understanding. This offers confidence in the reliability and interpretability of the model in Real-World production environments,” Lee further explained. “We believe that this methodology can be expanded further than the production of solar cells to a wide range of industrial processes.”
The proposed approach was described in the study “Bayesian Optimization-based approach for leaf resistance control in silicon waffles to automated solar cell production“Published in Material science in semiconductor processing.
This content is protected by copyright and may not be reused. If you want to work with us and reuse part of our content, please contact: editors@pv-magazine.com.
