AI 中文总结
本研究针对分布式孔径望远镜的共相挑战，提出基于CNN的焦平面图像对准误差估计方法，实现微米级快速粗对准，为未来高分辨率干涉系统的共相策略奠定基础。
AI 中文摘要
系外生命探测器（ELF）是一台35至50米级的斐索干涉望远镜，旨在通过直接成像探测系外行星的生物特征；小型系外生命探测器（SELF）是其3.5米原型机，用作该设施所需创新技术的测试平台。分布式孔径的精确共相是核心挑战，子孔径间的光程差必须保持在工作波长的极小分数范围内。本研究采用卷积神经网络（CNN）直接从焦平面图像估计对准误差，实现微米级的快速数据驱动粗对准。开发了一种基于自定义CNN的监督回归框架，用于建模子孔径对准误差与焦平面强度分布之间的非线性映射；训练和验证数据集由简化SELF系统的高保真光学模拟生成，采用高斯噪声增强评估鲁棒性，并对比评估多种现有CNN架构。所提方法重构了所有镜对的活塞和倾斜对准误差，证明可利用焦平面干涉图案进行对准估计；自定义CNN在估计性能、鲁棒性与计算效率间取得良好平衡，推理时间达毫秒级。研究发现性能与噪声鲁棒性存在权衡：噪声增强可提升低信噪比下的重构效果，但在近无噪声条件下性能下降。这些结果证明了分布式孔径望远镜采用数据驱动焦平面波前传感的可行性，为未来高分辨率干涉系统的自主、高效共相策略奠定了基础。
英文摘要
The ExoLife Finder (ELF), a 35--50-meter-class Fizeau interferometric telescope, was designed to detect biosignatures on exoplanets through direct imaging. The Small ExoLife Finder (SELF), a 3.5-meter prototype, serves as a testbed for the innovative technologies required for such a facility. A key challenge is the precise cophasing of its distributed aperture, where the optical path differences between subapertures must remain below a small fraction of the operating wavelength. This work investigates convolutional neural networks (CNNs) to estimate alignment errors directly from focal-plane images, enabling fast, data-driven coarse alignment to the few-micrometer level. A supervised regression framework was developed using a custom CNN to model the nonlinear mapping between subaperture misalignments and focal-plane intensity distributions. Training and validation datasets were generated from high-fidelity optical simulations of the simplified SELF system. Gaussian-noise augmentation was used to assess robustness, and several established CNN architectures were evaluated for comparison. The proposed approach reconstructed piston and tilt misalignments across all mirror pairs, demonstrating that focal-plane interference patterns can be exploited for alignment estimation. The custom CNN provided a favorable balance between estimation performance, robustness, and computational efficiency, with millisecond-scale inference times. A trade-off between performance and noise robustness was identified, with noise augmentation improving reconstruction at low signal-to-noise ratios but reduced performance under near noise-free conditions. These results demonstrate the feasibility of data-driven focal-plane wavefront sensing for distributed-aperture telescopes and provide a basis for autonomous, computationally efficient cophasing strategies for future high-resolution interferometric systems.
