ML-Based XPS Quantification Supported by Synthetic Dataset Generation

Abstract With growing interest in laboratory automation and high-throughput systems, the amount of generated experimental data is rapidly increasing while analysis methods still require many manual work hours from experts. This is prevalent in X-ray photoelectron spectroscopy (XPS), where quantification is a complex, time-consuming, and error-prone task. We therefore propose a neural network-based workflow to make this process more approachable. As training data availability ranges from insufficient to non-existent, our workflow creates a synthetic dataset containing XPS signals and corresponding area percentages based on binding energies supplied by the user. As a result, no previous measurements are needed. After training on the synthetic data, the neural network can predict area percentages of the known binding energies with high confidence. This workflow can therefore be adapted for XPS quantification tasks to filter significant data and supervise processes. Moreover, this enables non-experts to analyze spectra and can help experts to reduce focus on important spectra.