Deep Learning Enabled Ultrafast Optical Spectral Phase Retrieval via Unbalanced-Intensity Autocorrelation

Complete characterization of ultrafast optical fields is essential for deploying femtosecond pulse sources in integrated photonic systems, yet conventional techniques rely on spectrally resolved detection and iterative phase retrieval. Recently, phase-enabled nonlinear gating with unbalanced intensity (PENGUIN) showed that spectral phase information is preserved in an unbalanced nonlinear interferometric autocorrelation (IAC) enabling field retrieval without spectrally resolved detection. However, the original iterative reconstruction algorithm is only reliable for pulses near the Fourier-transform limit. To overcome this dispersion limitation, we introduce DeepPENGUIN, a deep-learning framework for direct spectral phase inference from unbalanced-intensity IAC signals in the frequency domain without iterative retrieval. The model is trained on large-scale synthetic data through physics-informed supervised learning with spectral and mathematical constraints, providing a robust solution for the one-dimensional phase retrieval problem for which no general inversion framework has been established. Validation using real-world few-cycle (8 fs) pulses demonstrates reliable phase retrieval over a large dispersion range from 0 to 200 fs². A quantitative noise robustness analysis shows performance comparable to ptychographic frequency-resolved optical gating across a broad range of noise levels. The feedforward inference architecture enables real-time pulse reconstruction with millisecond-level latency, providing a scalable route toward integrated ultrafast pulse diagnostics and feedback control.