Experimental investigation and prediction modelling of slicing speed and surface roughness during wafer slicing using WEDM

Abstract

To keep up with the rising demand for silicon solar cells in the photovoltaic sector, an alternative slicing method that can achieve high throughput with minimal waste is required. In recent research efforts, Wire electro-discharge machining (WEDM) has become the possible alternative method for slicing. The experimental investigation focuses on slicing monocrystalline silicon using the WEDM process with a brass wire electrode of 250 μm in diameter. The face-centered central composite design was employed for planning and conducting experiments. The investigational experiments were conducted with five different process parameters serving as inputs: peak current, wire tension, wire feed rate, pulse on and off time. The response parameter measured was the slicing speed and the surface roughness. Further, comparisons were made between different kernel functions in support vector regression (SVR) for the prediction modelling of slicing speed and surface roughness. The difficulty in prediction modelling can be attributed to the complexity of the WEDM process, which is caused by the involvement of various process parameters. The primary purpose of this work is to determine the best predictive kernel among the linear, polynomial, radial basis function (Rbf), and sigmoid kernel functions based on the experimental data. The predictive performance of different kernel functions was evaluated and compared. Grid search was used for the hyper tuning of the kernel parameters. The radial basis function produces R2 of 99.751% and 97.552%, MSE values of 0.00046 and 0.00079, RSME values of 0.0215 and 0.02814, MAE values of 0.01645 and 0.01894, and MAPE values of 1.2% and 0.9% for slicing speed and surface roughness. Support vector regression with radial basis function gives better results in comparison to other kernel functions, which concludes that support vector regression with radial basis function is well suited for the prediction of slicing speed and surface roughness.