Ultrafast Pulse Retrieval from Partial FROG Traces Using Implicit Diffusion Models

Abstract

Ultrashort laser pulses enable attosecond-scale measurements and drive breakthroughs across science and technology, but their routine use hinges on reliable pulse characterization. Frequency-Resolved Optical Gating (FROG) is a leading solution, forming a spectrogram by scanning the delay between two pulse replicas and recording the nonlinear signal spectrum. In online settings, however, dense delay-frequency scans are costly or impractical-especially for long pulses, wavelength regimes with limited spectrometer coverage (e.g., UV), or hardware with coarse resolution, yielding severely undersampled FROG traces. Existing reconstruction methods struggle in this regime-iterative algorithms are computationally heavy, convolutional networks blur fine structure, and sequence models are unstable when inputs are discontinuous or sparse. We present a generative diffusion framework tailored to recover ultrafast pulse intensity and phase from incomplete FROG measurements. Our model infers missing spectro-temporal content with high fidelity, enabling accurate retrieval from aggressively downsampled inputs. On a simulated benchmark of FROG-pulse pairs, the diffusion approach surpasses strong CNN and Seq2Seq baselines in accuracy and stability while remaining efficient enough for near real-time deployment.