摘要： The full characterization of ultrashort pulses is crucial for all their applications. Lately, the dispersion scan (d-scan) technique has been developed, which stands out by its simplicity [1, 2]. Therefore, it is a promising candidate for a low-maintenance pulse characterization technique. The reconstruction of the spectral phase from a d-scan measurement relies on optimization algorithms [3, 4]. Considering the fact that the spectral phase of the pulses is directly imprinted on the d-scan trace, the phase reconstruction can also be regarded as an image recognition task. During the last few years, artificial neural networks have shown excellent performance in different fields, e.g. image recognition [5]. Here we present the retrieval of spectral phases from d-scan traces using deep neural networks (DNN) that is more than 3000 times faster than conventional optimization algorithms (16 ms) [6]. Our DNN is an optimized variant of the DenseNet-BC architecture that has already shown great performance in image recognition tasks [7]. Training our shallow variant of the DenseNet takes only 12 h using a low cost GPU (NVIDIA GeForce GTX 1080 Ti). After training with up to 1.5 × 106 different, randomly generated traces, we can employ the network for reconstruction of the spectral phases from a d-scan traces. Fig. 1(a) and (b) show a randomly generated d-scan trace and its corresponding reconstruction, respectively. The spectral intensity (black curve), the original spectral phase (blue curve) and the retrieved phase (red dashed curve) are depicted in Fig. 1(c). Additionally, Fig. 1(d) compares the original and the reconstructed temporal pulse shapes (same color code). The consistency between the original and the retrieved phases and their corresponding temporal shapes shows the excellent performance of the DNN. Furthermore, to prove the fidelity of phase reconstruction from experimentally measured traces, we retrieved the spectral phase from a Ti:sapphire based amplifier system [8]. Compared to a differential evolution optimization algorithm, we observe identical full width at half maximum pulse durations, as well as nearly identical temporal pulse shapes. We believe that DNNs combined with single-shot d-scan setups [9] enable video-rate measurements with an inherent error estimation due to the redundancy in the data.