<p>Joint inversion of gravity and magnetic data is essential for subsurface characterization, yet traditional methods require accurate geological models—often unknown for deep or complex structures. Although networks like U-Net have been widely shown to outperform traditional inversion, they still struggle to accurately invert complex, decoupled physical property distributions due to the absence of a decoupling mechanism, rigid mapping without perturbation strategies. To address these issues, we propose PFInvNet, a 2D double branch multi-channel convolutional encoder-decoder network to enable efficient 3D joint inversion. It features a double-branch structure, a self-attention module (SAM), and a feature fusion module (FFM) to jointly recover density and susceptibility while respecting their physical independence and observational coupling. Tested on diverse unseen synthetic models with fully decoupled property assignments, all PFInvNet variants—including those without SAM or FFM—consistently outperform U-Net under identical training settings, confirming the robustness of the base architecture. The full model achieves the highest accuracy and structural fidelity, with reduced background noise and improved resolution of adjacent bodies. Validation on real data from the Victoria Land Basin yields geologically plausible 3D property distributions, demonstrating PFInvNet’s practical value even without detailed prior models.</p>