Virtual photon diagnostics at SwissFEL for real-time spectral analysis

We present a method for real-time spectral analysis of SASE X-ray FEL spectra using a machine learning (ML) model. The model detects and counts SASE spikes, even when their widths fall below the spectrometer resolution. For a typical 20-30 fs SASE pulse in the hard X-ray regime, the individual spike width is below the intrinsic resolution of bent crystal online spectrometers, making it impossible to count spikes using standard methods accurately. By training an ML model on synthetic SASE spectra with a known number of spikes, convolved with the intrinsic spectrometer resolution, the virtual diagnostic provides an accurate spike count despite resolution limitations. The model provides the number of spikes in the single-shot spectra in real time to the data system of SwissFEL, serving as a key online monitor for likely changes (both desired and unwanted) to the FEL pulse duration. Trained on SwissFEL datasets, this approach enables fast, accurate spike counting measurements. Benchmarked against the electron streaking cavity, the method provides operators with a robust measurement of X-ray pulse duration without a complex direct measurement of the pulse duration.