Rigaku logo
Semiconductor
Read Article

Machine Learning and Application to Spectral Analysis on TXRF Spectrometry

Makoto Doi and Shinya Kikuta

Summer 2021 Volume 37, No. 2 , 26-32

Total Reflection X-ray Fluorescence (TXRF) analysis is a non-destructive and surface-sensitive analysis method using X-rays, in which incident X-rays are irradiated on a sample at an extremely low grazing angle (about 0.1°) and the fluorescent X-rays from the sample generated by the incident X-rays are measured with extremely low background because of the total reflection characteristics of the incident X-rays. TXRF analysis does not require special sample preparation for flat samples. Because of this, TXRF analysis has been widely used for the evaluation of contamination on wafers in semiconductor manufacturing processes(3) as well as in industrial and environmental analysis. Contamination control in semiconductor manufacturing processes becomes more rigorous every year.

Recently, Artificial Intelligence (AI) technologies have developed rapidly along with progress in computer hardware, software and software libraries to deal with big data. One main benefit of AI is that it automatically extracts and analyzes unique and notable characteristics from a huge amount of data. In the field of image processing, particularly, image recognition—for example, handwritten character recognition—has been actively researched and many results—such as super-resolution techniques that convert low-resolution images to high-resolution ones—have been achieved. Although there are many cases where AI is used for image processing, it seems that there are few cases where AI technologies are applied to one-dimensional spectrum analysis instead of to a two-dimensional image. Therefore, in this paper, we applied the machine learning method to the data processing of TXRF analysis and introduce the results, especially on the quantification of contaminations on wafers from the spectrum obtained by short-time measurements.

Highlights

  • Machine learning significantly improves the detection of weak TXRF spectral peaks during short measurement times, reducing both false-positive and false-negative contamination detections.
  • A convolutional neural network trained on short-duration spectra can produce results that closely match conventional analysis from much longer measurements, improving throughput without sacrificing accuracy.
  • Fine-tuning a pretrained machine learning model with approximately 1,000 instrument-specific samples enables effective transfer to a different TXRF system while maintaining strong analytical performance.

Summary

Total Reflection X-ray Fluorescence (TXRF) is widely used for detecting trace surface contamination on semiconductor wafers, where rapid measurements are essential for maintaining manufacturing throughput. Conventional peak-fitting methods perform well when spectral peaks are strong but become less reliable for weak signals collected during short acquisition times, leading to missed contaminants or false detections.

A machine learning approach based on a convolutional neural network addresses this challenge by learning the relationship between five-second TXRF spectra and reference results obtained from sixty-second measurements. Instead of relying solely on conventional spectral deconvolution, the model identifies subtle spectral features that distinguish true elemental signals from statistical noise. Across a broad range of contamination elements, the machine learning model improved agreement with long-duration measurements and substantially reduced false-positive and false-negative results, particularly for low-intensity peaks.

The work also demonstrates that pretrained models can be adapted to different TXRF instruments. While a model transferred directly to another instrument produced mixed results, fine-tuning with approximately 1,000 instrument-specific training samples restored high correlation with conventional long-measurement analyses while preserving the reduction in false detections. The findings suggest that combining machine learning with traditional peak fitting offers an effective strategy for improving the speed, sensitivity, and reliability of TXRF contamination analysis in semiconductor manufacturing, particularly when rapid wafer mapping is required.

Frequently asked questions

Short acquisition times produce fewer detected X-ray photons, making weak elemental peaks difficult to distinguish from statistical noise. Conventional peak-fitting methods may incorrectly classify these weak signals, resulting in false-positive detections of contamination or false-negative failures to detect trace contaminants. These errors become especially significant in high-throughput applications where measurements may last only five to ten seconds.

Machine learning can learn complex relationships between measured spectra and known elemental concentrations by training on large collections of labeled data. Rather than relying exclusively on mathematical peak fitting, a neural network recognizes subtle spectral patterns associated with real elemental signals, improving quantification accuracy and reducing incorrect detections in noisy, short-duration measurements.

A false positive may trigger unnecessary cleaning, process adjustments, or rejection of acceptable wafers, increasing manufacturing costs. A false negative allows actual contamination to go undetected, potentially reducing device yield or reliability. Minimizing both types of errors improves process control while maintaining production efficiency.

Longer measurement times produce spectra with higher counting statistics and more reliable elemental quantification. These higher-confidence results serve as reference values during supervised learning, allowing the model to learn how to predict accurate elemental intensities from much shorter measurements that contain significantly more statistical noise.

Yes, but performance depends on how similar the instruments and measurement conditions are. A model transferred directly may experience reduced accuracy because of differences in detector response, calibration, or instrument characteristics. Fine-tuning the pretrained model with a relatively small set of data from the target instrument substantially improves both quantitative accuracy and detection reliability.

Machine learning excels at identifying weak peaks and reducing false detections in noisy spectra, while conventional peak-fitting methods remain highly effective for accurately quantifying strong peaks. Combining the two approaches leverages the strengths of each method, providing improved sensitivity for trace contaminants without sacrificing quantitative performance for higher-concentration elements.

The greatest benefits are realized in applications requiring rapid, high-throughput measurements with high sensitivity, such as full-wafer contamination mapping in semiconductor manufacturing. These environments demand both fast analysis and reliable detection of extremely small amounts of surface contamination to maintain manufacturing yield and process quality.

Recommended products

Subscribe to Understanding Semiconductors newsletter

Connect with semiconductor leaders in metrology, process, and analytics to solve your biggest metrology challenges.

Subscribe

Contact Us

Whether you are interested in getting a quote, want a demo, need technical support, or simply have a question, we're here to help.

Contact Sales
Contact Service
Office Locations