Abstract
Ultra-short laser pulses with femtosecond to attosecond pulse duration are the shortest systematic events humans can create. Characterization (amplitude and phase) of these pulses is a key ingredient in ultrafast science, e.g., exploring chemical reactions and electronic phase transitions. Here, we propose and demonstrate, numerically and experimentally, the first deep neural network technique to reconstruct ultra-short optical pulses. We anticipate that this approach will extend the range of ultrashort laser pulses that can be characterized, e.g., enabling to diagnose very weak attosecond pulses.
Introduction
The ability to generate and measure ultra-short laser pulses is fundamental to modern ultrafast science. These pulses, lasting only femtoseconds ($10^{-15}$ s) or attoseconds ($10^{-18}$ s), allow researchers to observe atomic and molecular dynamics in real-time. However, characterizing these pulses presents a unique challenge: measuring the shortest event in existence requires an even shorter "gate," which does not exist.
The FROG Challenge
To circumvent this, scientists employ Frequency-Resolved Optical Gating (FROG). In a typical Second Harmonic Generation (SHG) FROG setup, a pulse is split and recombined in a non-linear crystal with a variable time delay. The resulting spectrogram, or "FROG trace," contains the information needed to reconstruct the pulse's temporal amplitude and phase.
The reconstruction process is a non-linear phase-retrieval problem. Traditional iterative algorithms, such as the Principal Components Generalized Projection Algorithm (PCGPA), are computationally expensive and often fail in high-noise environments. This limitation is particularly acute for weak attosecond pulses, where the signal-to-noise ratio (SNR) is frequently below the threshold for reliable recovery.
DeepFROG: A Physics-Informed Deep Learning Approach
We introduce DeepFROG, a framework that utilizes Convolutional Neural Networks (CNNs) to perform near-instantaneous pulse reconstruction. DeepFROG is designed to bridge the gap between theoretical simulation and experimental reality through a multi-stage training process.
Architecture and Training Pipeline
The core of DeepFROG is a CNN (utilizing "Multires" or DenseNet architectures) trained to map 2D traces to 1D pulse fields. The training involves two primary phases:
- Supervised Learning: The network is trained on 60,000 simulated pulse/trace pairs. Crucially, Additive White Gaussian Noise (AWGN) is injected during this phase, teaching the network to inherently filter noise and handle low-SNR signals.
- Unsupervised Refinement: To adapt to experimental data where ground truth is unavailable, we employ a "FROG Net"—a fixed, differentiable layer that implements the physical FROG measurement equation. By comparing the trace generated from the CNN's prediction to the actual experimental measurement, the network can be refined without explicit labels.
Results and Discussion
DeepFROG demonstrates significant improvements over traditional methods. In simulations, it maintains high accuracy at SNR levels below 20dB, where iterative solvers typically break down. When applied to experimental data from a Ti-Sapphire laser, the hybrid training approach achieved the lowest reconstruction error among all tested methods. Furthermore, once trained, the inference is nearly instantaneous, enabling real-time pulse characterization.
Conclusion
DeepFROG represents a paradigm shift in ultrafast pulse characterization. By integrating physical constraints into a deep learning framework, we have developed a robust, high-speed tool capable of reconstructing pulses in regimes previously considered inaccessible. This approach paves the way for the characterization of extremely weak attosecond pulses, opening new frontiers in the study of ultrafast phenomena.
Summary updated