High-quality beam transmission is an important guarantee for the stable operation of precision optical systems. As a key component for suppressing high-frequency noise of laser beams and improving beam shaping quality [1,2], spatial filters are widely used in interferometry [3], polarization imaging [4], laser communication [5] and high-energy laser systems [6]. The performance of spatial filtering systems depends to a large extent on the alignment accuracy of the lenses. When the lens has small six-degree-of-freedom (6-DOF) pose errors such as eccentricity, tilt, or defocus in space, the energy distribution and wavefront characteristics of the light spot will change nonlinearly, resulting in a decrease in system resolution and a reduction in optical power coupling efficiency. Therefore, achieving high-precision automatic alignment of spatial filtering lenses is of great significance for improving the performance of optical systems.

At present, the alignment of spatial filter lenses still mainly rely on manual operation, usually by observing the shape of the light spot or interference fringes for fine adjustment. This method is not only inefficient and has poor repeatability, but also difficult to meet the requirements of high-precision systems for alignment stability and consistency. In order to improve alignment accuracy and efficiency, Zhang et al. [7] designed a new type of parallel adjustment mechanism for filter lenses, which can realize five-dimensional precision adjustment of the lens in the X-Y-Z axis translation and pitch and yaw directions, providing higher mechanical adjustment accuracy for the spatial filtering process. Gao et al. [8] developed an automatic collimation system for the SGII-Up laser facility, achieving far-field pointing accuracy better than 3 μrad and near-field centering accuracy less than 1 mm. In recent years, machine learning has also begun to play a transformative role in the precision control of high-power lasers; for example, recent research has successfully demonstrated the use of predictive machine learning to achieve real-time pointing stabilization of high-power lasers, proving its great potential in overcoming the limitations of traditional closed-loop feedback delay [9].Peñas et al. [10] reported a multi-target automatic alignment program for high repetition rate laser-plasma proton acceleration, which can accommodate more than 5000 targets and operate at a rate of up to 10 Hz. In addition to mechanical and system-level solutions, model-based optimization algorithms are another technical path. Parinam et al. [11] used an improved genetic algorithm to optimize high-transmittance filters, with approximately 99% transmittance in the 525–575 nm wavelength region and no drop-off in between being the main performance requirement. However, whether it is precision mechanics, complex systems or model algorithms, when facing small pose errors in multi-degree-of-freedom coupling of lenses, there are still challenges such as complex alignment processes, reliance on expensive sensors or difficulty in learning complex nonlinear relationships.

In recent years, with the development of deep learning [12,13], some studies have combined neural networks with optical systems to realize optical polarization state detection, waveform denoising and computational imaging. It is worth noting that the current algorithmic paradigm in the field of optics is rapidly shifting from purely data-driven models to physics-enhanced methods. These methods, by incorporating prior knowledge of optical physics, effectively alleviate the challenges of acquiring large-scale data and significantly improve the interpretability of the models [14]. For example, Xiong et al. [15] used convolutional neural networks to establish the mapping relationship between speckle parameters and Stokes parameters, and constructed a simple, low-cost, portable SOP measurement system based on multimode fiber speckle. Chen et al. [16] introduced a U-shaped convolutional neural network, which integrates multi-scale feature extraction capability, attention mechanism and long short-term memory unit, effectively promoting real-time denoising of various shaped pulses. Shen et al. [17] used deep learning to train multi-layer spatial diffraction structures to convert the phase distribution of objects at different axial positions into intensity patterns of independent wavelength channels, thereby realizing the reconstruction of quantitative phase distribution of multiple planes using only intensity image sensors.

In particular, convolutional neural networks [18,19] have shown strong capabilities in image feature extraction and pattern recognition. In the field of optics, CNNs have been used in tasks such as spot recognition [20], phase retrieval [21], wavefront aberration estimation [22] and light field reconstruction [23]. By automatically extracting features from a large number of spot images, nonlinear mapping of parameters of complex optical systems has been realized. Recent research shows that training neural networks with high-fidelity physical simulation data can significantly improve the robustness of models in complex light fields [24]. For example, Li [25] combined a lightweight convolutional neural network (CNN) with a Kalman filter. This system can accurately detect laser spots and adjust the beam direction through a closed-loop feedback mechanism. Li et al. [26] proposed a phase retrieval network based on two working modes of convolutional neural networks. This network can not only obtain a persistent model through the training dataset, but also construct a special loss function under the Gaussian measurement model. Nishizaki et al. [27] used convolutional neural networks to directly estimate the Zernike coefficients of the aberration wavefront from a single intensity image. The proposed deep learning wavefront sensor can be trained to estimate the wavefront aberrations excited by point sources or even extended sources. In the field of lens alignment, the Slor team [28] explored a complementary reverse design method based on deep learning, achieving an average absolute error of 0.031 mm for lateral translation and an average absolute error of 0.011° for tilting.

These studies demonstrate that the CNN-based intelligent adjustment method for spatial filtering lenses has great application potential and can provide a new technical path for the automation and intelligence of complex optical systems.

In summary, this paper presents a CNN-based intelligent alignment method for spatial filtering lenses. By constructing a spot image dataset incorporating 6-DOF micro-pose perturbations, a dual-channel convolutional network structure is designed to jointly input spatial and frequency domain information, enabling simultaneous prediction and evaluation of the lens's six-axis deviations. The proposed approach achieves end-to-end regression mapping from spot images to pose parameters and demonstrates its feasibility and accuracy through experimental validation. The results provide both theoretical support and practical guidance for the automated and high-precision alignment of spatial filtering lenses in high-power laser systems.