Data analysis in angle-resolved photoemission spectroscopy (ARPES) and time-resolved ARPES usually employs derivative-based methods, such as second derivative and maximum curvature, to enhance spectral visibility. Despite their widespread use, these methods are prone to noise amplification and the introduction of artificial features. While deep learning offers a promising alternative, its effectiveness typically depends on large-scale experimental datasets. Here, we propose a Sim2Real deep learning approach that enables effective denoising without the requirement of extensive experimental data. By implementing a gradient attention U-net model, the signal-to-noise ratio is improved significantly across multiple datasets. Taking the parabolic dispersion bands for nearly free electron and the polaron features for correlated-electron phenomena as examples, our approach exhibits good performance in processing ARPES spectra, in comparison with the derivative-based methods. Moreover, our approach enables up to an $\sim 500$-fold reduction in data acquisition time—a critical advantage for time-consuming ultrafast time-/phase-resolved ARPES experiments. These capabilities underscore the potential of deep learning for quantitative, high-throughput photoemission and other spectroscopies.

Close

PrevNext

View figure in article

• Surfaces
• Angle-resolved photoemission spectroscopy
• Data analysis
• Deep learning