High-Throughput Quantification of XPS Spectra using Explainable 1D-Convolutional Neural Network

X-ray photoelectron spectroscopy serves as a cornerstone of materials analysis, yet its manual interpretation remains a major bottleneck in high-throughput research due to human subjectivity. To address this, we propose an automated quantitative analysis framework based on an explainable 1-dimensional convolutional neural network optimized for rapid spectral screening. To ensure physical grounding, we established a foundational library by experimentally measuring 76 pure elements individually from Li to Tl, creating an augmented training set of 320,000 spectra. Our model achieved an expert-level accuracy of 90.6% and exhibited stable performance at a signal-to-noise ratio as low as 3.8, enabling reliable high-throughput characterization with reduced acquisition times. A pivotal contribution of this work is the transparent elucidation of the AI's decision-making pathways using a Gaussian probing-based Feature Importance Spectrum (FIS) technique. By visualizing the model's focus along the binding energy axis, we confirmed that the AI identifies elements based on objective spectroscopic physical reality rather than mere numerical artifacts. This physical decoding enabled us to establish a four-class strategic classification system based on spectral overlap complexity, providing the first systematic interpretational guidelines across the periodic table for AI-driven XPS analysis. This framework offers a robust methodology to minimize human error and accelerate materials discovery. Beyond XPS, this approach serves as a universal methodology applicable to other spectroscopic domains such as Raman and FTIR, ensuring high reliability in data-driven research.