This paper introduces a novel gradient-based design framework for large-area, high-numerical-aperture, freeform metasurfaces. The framework combines the simplicity and fabricability of conventional design paradigms with the extended capabilities of freeform optimization. By leveraging nonlocal interactions between simply shaped nanostructures placed on an irregular lattice, the design can produce high-order hybridized modes that support customizable large-angle scattering profiles. The authors demonstrate the design and experimental validation of multifunctional super-dispersive metalenses and high numerical aperture radial metalenses capable of diffraction-limited focusing and generating donut-shaped point spread functions. These concepts are anticipated to have applications in super-resolution microscopy, particle trapping, additive manufacturing, and metrology. The design approach involves subdividing the wavefront profile into wavelength-scale "super-pixel" regions, each optimized for specific scattering angles, phases, and polarizations. The method ensures computational tractability and scalability while incorporating hard geometrical constraints through reparameterization. The results show significant improvements in focusing performance compared to conventional designs, with applications in various scientific and technological domains.This paper introduces a novel gradient-based design framework for large-area, high-numerical-aperture, freeform metasurfaces. The framework combines the simplicity and fabricability of conventional design paradigms with the extended capabilities of freeform optimization. By leveraging nonlocal interactions between simply shaped nanostructures placed on an irregular lattice, the design can produce high-order hybridized modes that support customizable large-angle scattering profiles. The authors demonstrate the design and experimental validation of multifunctional super-dispersive metalenses and high numerical aperture radial metalenses capable of diffraction-limited focusing and generating donut-shaped point spread functions. These concepts are anticipated to have applications in super-resolution microscopy, particle trapping, additive manufacturing, and metrology. The design approach involves subdividing the wavefront profile into wavelength-scale "super-pixel" regions, each optimized for specific scattering angles, phases, and polarizations. The method ensures computational tractability and scalability while incorporating hard geometrical constraints through reparameterization. The results show significant improvements in focusing performance compared to conventional designs, with applications in various scientific and technological domains.