A novel computational methodology based on statistical learning multiobjective optimization is developed to optimize large-scale achromatic 3D metalenses in the visible regime. The optimized lens has a numerical aperture of 0.56 and an average focusing efficiency of 45%. The metasurface is a nanostructured interface with a subwavelength thickness which manipulates light through spatially arranged meta-atoms. These meta-atoms ensure adequate control of the light attributes by altering the phase, amplitude and polarization of the incoming wavefront [1-2]. Due to the versatility of the metasurface concept, peculiar and outstanding applications have been demonstrated recently [3-4]. Yet, the precise engineering of the subwavelength meta-atoms, in particular for extremely challenging applications, is therefore crucial for improving the performance. As a result, numerous inverse design methodologies have recently been adopted to further broaden the potential of a metasurface design, in particular for single-objective based devices [5]. Nevertheless, the next decade is expected to witness a notable increase in the field of multifunctional metasurfaces owing to the increasing demand in the design of flat optical devices with a diverse variety of functionalities. This makes it a mandatory step for the next metasurface generation to leverage robust inverse design techniques, which are capable of optimizing metasurfaces with multiple functionalities. However, this is not an easily manageable task, since the optimization of multifunctional metasurfaces requires a solution to an optimization problem with different conflicting objectives and potentially entails a large parameter space.