PB-ND: A Physics-Informed Neural Network Scheme for Astronomical Image Deconvolution

PB-ND is a physics-informed neural network (PINN) scheme designed specifically for astronomical observation data. Its core lies in embedding optical physics equations into the loss function to achieve high-quality deconvolution reconstruction of JWST and Pan-STARRS data. It addresses the "hallucination" or conservatism issues of traditional deconvolution methods, optimizing image quality while respecting physical laws, and has significant scientific value and application prospects.

Deep space observation faces issues such as photon noise, atmospheric turbulence (for ground-based observations), and PSF blurring. Traditional deconvolution methods are in a dilemma: being too aggressive leads to hallucinations, while being too conservative fails to tap potential. Physics-informed neural networks (PINNs) are an important advancement in scientific machine learning. By simultaneously fitting data and satisfying physical laws, they perform excellently in astronomical scenarios with sparse data or strong noise.

PB-ND uses a modified U-Net as its backbone network. The core innovation is the Physics-Informed Loss Engine (hybrid L1/MSE loss + PSF wavelength convolution term), which ensures the output complies with optical diffraction laws; Hyperspectral Calibration (percentile Z-scaling + arcsinh stretching) handles the dynamic range of infrared data; Sliding Window Stitching solves the memory limit for large-scale data and supports processing on ordinary CPUs.

PB-ND's effectiveness is verified through metrics such as SSIM and PSNR. Key metric: A 45-second exposure image processed by PB-ND is equivalent to a 6.8-minute exposure (9x improvement). Comparison images of the Orion Bar show that the reconstructed results have improved detail sharpness and signal-to-noise ratio in faint areas, without the ringing artifacts of traditional methods.

PB-ND is built on PyTorch. It uses Astropy to process FITS files, NumPy for tensor operations, and Scikit-Image to calculate quality metrics. The loss function is exquisitely designed: dynamic masking of saturated star cores prevents gradient explosion, and strong constraints on local structural similarity avoid sacrificing details in faint areas—this is the key to avoiding "hallucinations".

PB-ND represents a new paradigm of "physics-constrained machine learning" and has broad prospects in fields such as astronomy and medical imaging. For JWST/Pan-STARRS: it enhances the value of archived data (reprocessing historical observations) and enables detection of fainter transient events; for time-domain astronomy, the equivalent exposure improvement is of great significance.

PB-ND is released as open-source with clear code and complete documentation, supporting FITS format and seamless integration with existing workflows. It provides a reproducible tool for the astronomy community and demonstrates to ML researchers the application mode of PINNs in scientific problems (e.g., converting physical equations into differential losses, balancing data and physics).

PB-ND is an excellent case of the intersection between AI and basic science. It replaces data piling with physical priors, embodying the "less is more" concept. For AI for Science developers, it is a high-quality project worth studying, bringing profound inspiration to the field of scientific computing.