基于深度學習的分孔徑相干合成系統相位控制方法

相干合成技術是突破單路激光亮度提升限制的重要技術途徑,已成為激光技術領域的研究前沿和熱點。按照孔徑填充方法的不同,相干合成可以分為共孔徑和分孔徑合成兩類。分孔徑相干合成通過相位控制實現各子光束之間的波前匹配,有效地增大了發射口徑、壓縮了遠場發散角,實現了亮度的提升。

在分孔徑相干合成系統中,各路激光的動態相位噪聲嚴重影響合成系統的效率、能量集中度和亮度,是限制合成系統向高功率、大陣元數目拓展的關鍵因素之一。為了克服動態相位噪聲的影響,國內外科研人員相繼提出了外差法、多頻抖動法、單頻抖動法、隨機并行梯度下降算法等多種相位控制方法,并有效應用到不同類型的相干合成系統中。然而,上述相位控制方法存在一個共性難題:隨著合成路數的拓展和相位噪聲的增強,系統控制帶寬會出現不同程度的下降,進而導致鎖相殘差增加、合成效率下降。

國防科技大學前沿交叉學科學院課題組首次將深度學習應用于分孔徑相干合成系統,為解決上述難題提供了新的參考思路:構建一個卷積神經網絡,在經過預先訓練后,該網絡便可以準確反映相位誤差與可測量的合成光束光強分布之間的關系,進而直接補償相位噪聲。傳統相位控制方法通常從合成光束的遠場光強信息中提取相位控制信號;然而,深度學習的引入帶來了新的技術挑戰——合成光束的遠場光強分布與陣列單元的相對相位之間并沒有一一對應的關系,可能會因數據沖突導致網絡失效。為了克服這個困難,該課題組提出在非焦平面上訓練、以有效提取相位信息的方法,結果表明,在非焦平面處訓練的卷積神經網絡可以更準確地反映發射面光束陣列的相位分布,進而實現了動態相位噪聲的高效精確補償。

在此基礎上,課題組通過7單元和19單元正六邊形陣列相干合成系統對基于深度學習的相位控制方法進行了測試。利用合成光束的關鍵指標(包括斯特列爾比、條紋對比度和桶中功率)評估相位控制性能,驗證了基于深度學習的相位控制方法的可行性和可拓展性。隨著陣列單元數量的拓展,這種方法不會導致卷積神經網絡的計算耗時和系統的復雜性增加,并且與經典的優化算法和抖動技術兼容。該研究結果發表在High Power Laser Science and Engineering 2019年第7卷第4期上(Tianyue Hou, Yi An, Qi Chang, Pengfei Ma, Jun Li, Dong Zhi, Liangjin Huang, Rongtao Su, Jian Wu, Yanxing Ma, Pu Zhou. Deep-learning-based phase control method for tiled aperture coherent beam combining systems[J]. High Power Laser Science and Engineering, 2019, 7(4): e59)。

該課題組周樸研究員表示:“我們將深度學習引入分孔徑相干合成系統,驗證了該方法的有效性和可拓展性,為解決大功率、大陣元相干合成系統的復雜控制難題提供了新的思路。”

這項工作中提出的方法為相干合成系統中相位控制帶寬隨陣列單元數量拓展而降低這一難題提供了新的解決方案。后續將進一步瞄準樣本數據采集、網絡結構設計和多維光場信息挖掘等方面對該控制方法進行綜合優化,力爭將該方法應用到大功率、超大陣元相干合成系統中。

實施基于深度學習相位控制方法的相干合成系統示意圖

Deep-learning-based phase control method for tiled aperture coherent beam combining systems

Coherent beam combining (CBC) technology is an important technical approach to break through the brightness limitation of a single laser beam, and has become a frontier and hotspot of laser technology research. According to the aperture filling method, CBC can be classified into two categories: filled aperture combining and tiled aperture combining. Tiled aperture CBC achieves wavefront matching among beamlets through phase control, thus efficiently increases the emission aperture, compresses the far-field divergence angle, and improves the brightness.

In tiled aperture CBC system, the dynamic phase noise of each laser beam seriously affects the efficiency, energy concentration, and brightness of the combining system, and it is one of the key factors that limits the development of the combining system to high power and large array elements. In order to overcome the impact of dynamic phase noise, researchers have successively proposed various phase control methods, such as heterodyne detection, multi-dithering technique, single-frequency dithering technique, and stochastic parallel gradient descent algorithm, which have been effectively implemented to different types of CBC system. However, the above-mentioned phase control methods have a common problem: with the expansion of the combining channels and the enhancement of phase noise, the system control bandwidth would decrease to some degree, which leads to an increase in phase-locked residuals and a decrease in combining efficiency.

A research group from College of Advanced Interdisciplinary Studies, National University of Defense Technology has incorporated deep learning (DL) into tiled-aperture CBC systems for the first time, providing a new reference on solving the above-mentioned problem. The authors have shown that constructing a convolutional neural network (CNN), which can accurately reflect the relationship between the phase error and the measurable intensity distribution of the combined beam after pre-training, and then the phase noise could be directly compensated. Conventional phase control methods usually extract the phase control signals from the far-field intensity information of the combined beam. However, the introduction of DL brings a new technical challenge-there is no one-to-one correspondence between the far-field intensity distribution of the combined beam and the relative phases of the array elements, which may cause network failure due to data confliction. To overcome this challenge, the authors propose a method for training the CNN at the non-focal-plane to effectively extract phase information. The results show that the CNN trained at the non-focal-plane could reflect the phase distribution of the beam array at the source plane more accurately, and moreover, efficient and accurate compensation of dynamic phase noise could be achieved.

Furthermore, the DL-based phase control method has been tested by the CBC systems with 7-element and 19-element hexagonal arrays. Using key metrics (including Strehl ratio, fringe contrast and power in the bucket) of the combined beams to evaluate the phase control performance, the researchers have demonstrated the feasibility and extensibility of the DL-based phase control method. As the number of array elements increases, such a direct phase control method would not cause the increases in the computing time of CNN and the complexity of the CBC system, and it is compatible with classical optimization algorithms and dithering techniques. The research results are published in High Power Laser Science and Engineering, Vol. 7, Issue 4, 2019 (Tianyue Hou, Yi An, Qi Chang, Pengfei Ma, Jun Li, Dong Zhi, Liangjin Huang, Rongtao Su, Jian Wu, Yanxing Ma, Pu Zhou. Deep-learning-based phase control method for tiled aperture coherent beam combining systems[J]. High Power Laser Science and Engineering, 2019, 7(4): e59).

"We have incorporated DL into tiled aperture CBC systems and verified the effectiveness and extensibility of the method, which provides a new idea for solving the complex control problem of high-power, large-array CBC systems. " said Professor Pu Zhou from the research grup.

The method proposed in this work provides a new solution to the problem that in the CBC system, the phase control bandwidth decreases as the number of array elements increases. Follow-up will further aim at sample data collection, network structure design and multi-dimensional light field information mining to comprehensively optimize the control method, and strive to implement this method to high-power, ultra-large array CBC systems.

Schematic of the deep-learning-based phase control method implemented coherent beam combining system.