Improvement of sensitivity in total reflection x-ray fluorescence spectrometry by machine learning

Abstract

Total reflection X-ray fluorescence (TXRF) spectrometry is widely used in semiconductor process control, enabling the detection of trace metal contamination on the surface of Si wafers. We are developing TXRF tools with higher sensitivity and better performance as contamination control becomes more critical. This is because the processes are more complicated by the demands for the miniaturization and high performance of semiconductor devices. In this work, we applied machine learning to TXRF. We prepared a training data set of 9000 TXRF measurements. We built a neural network model with one convolution layer and four hidden layers, and trained it to relate TXRF spectra to quantitative elemental composition. After training the neural network, we analyzed new TXRF data (independent from the training data) and obtained results which the neural network model inferred. As a result, the machine learning technique detected and analyzed the elements which were not detected by the conventional peak fitting methods in shorter measurement times. The results from the neural network were comparable to longer measurement times. We also discussed the possibility of achieving higher sensitivity by machine learning techniques.