#3rd place- Multiscale Aerodynamic Resolution Invariant Operator.

We present MARIO (Multiscale Aerodynamic Resolution Invariant Operator), a conditional neural field approach tailored for aerodynamic predictions. MARIO employs multiple input Fourier feature embeddings at different scales to optimize the reconstruction accuracy across the different frequency components, especially for multi-output predictions. The architecture is conditioned through a hypernetwork and FiLM modulation on the free-stream conditions and airfoil geometry, parameterized through thickness and camber distributions. We enhance the traditional coordinate and signed distance function inputs with a continuous normal vector field on and off the airfoil surface aiming to provide a translation invariant frame of reference, augmenting the implicit distance information. Additionally, a boundary layer mask is used in order to improve the network reconstruction performance close to the airfoil surface, where sharp variations of velocity are present. This has the effect of improving the integral drag accuracy. The method is inherently mesh-agnostic, enabling training on low-resolution data while performing super-resolution at inference time. MARIO demonstrates exceptional computational efficiency and speedup compared to the baseline solver, making it particularly suitable for industrial aerodynamic simulations.