CoFINN embeds conservation physics into CNNs, cutting drag errors by 34%

CoFINN (Conservation Flux Informed Neural Networks) is a new physics-informed deep learning framework developed by researchers Adnan Harun Doğan, Mert Deniz, Hande Alemdar, and Özgür Uğraş Baran. It targets compressible flow fields governed by conservation laws, a common challenge in aerospace and mechanical engineering. Unlike conventional data-driven convolutional neural networks that optimize pixel-wise similarity, CoFINN integrates finite-volume conservation physics directly into the training process. This contrasts with classical physics-informed neural networks that enforce differential-equation residuals at collocation points via automatic differentiation; instead, CoFINN adopts a finite-volume perspective consistent with modern CFD methodology.

The framework interprets CNN output fields as structured computational grids where each pixel is a finite-volume cell, enforcing conservation consistency through sophisticated numerical flux calculations. Tested on transonic flow prediction around airfoils at Mach 0.7 and Reynolds number 6×10^6—including challenging shock wave conditions and high angles of attack—CoFINN significantly improved aerodynamic force predictions. It reduced drag prediction error by up to 34% at extreme angles of attack and by approximately 15% on average across the full test set. Improvements were especially pronounced in limited-data regimes, where the conservation-based loss acts as an effective physical regularizer. The approach is architecture-agnostic and extensible to broader classes of conservation-law-governed physical systems, maintaining the computational efficiency of CNN surrogates while greatly improving physical consistency.