From Image Deconvolution to Spectral Fitting: Machine Learning in the EOVSA Pipeline

Yuqian Wei1, Bin Chen1, Sijie Yu1 1: New Jersey Institute of Tech DASH 2025 From Image Deconvolution to Spectral Fitting Machine learning in EOVSA pipeline Learned Fourier-Aware Imaging •Sparse short baseline coverage at high frequencies" •(Hard to recover large scale features)" •Expensive calculation Model Structure A: forward operator (Imag->Vis) A⁻¹ : backward operator (Vis->Imag) Input Channels Core model Loss Functions Reconstructed Image I_recon = A⁻¹(V_sim) Multi-Channel Input [dirty, ΔI] Visibility Difference ΔV In Image Domain ΔI = A⁻¹(V_sim −A(I_recon)) Auxiliary (PSF, weights, …) Δ1 = U1([x0, y0=0]) →y1 =y0 +Δ1 [ResNet +U-Net]*6 Clean Image Î_clean =f_θ(X) Δi = Ui([x0, yi-1]) →yi =yi-1 +Δi... 6 * U-Net Intensity Loss (flux weighted) α·L_intensity Fine Structure Loss δ·L_structure Multi-scale Loss λ·L_MS ... Backpropagation &Optimization θ* = arg min_θL_total (Adam optimizer) Spectral Resolution: Single spectral window(s) Dirty Img Rec Img Model Img Visibility Domain (physics constrain) Dirty Img FFT Rec Img FFT Model Img FFT Spectral fitting Pygsfit: Y.Wei et al Gsfit: G.Nita et al Chen et al., 2021 Global minimization & Sample the posterior(uncertainty) •neural density estimator (SNPE-C normalizing flow) to learn q(θ|x) and infer posterior for the observed spectrum Global Minimization •neural density estimator VS MCMC