Developing differentiable frameworks for inverse design of freeform optical systems including lens-reflector combinations. Key contributions include: (1) implicit differentiation through ray-surface intersection using the implicit function theorem, circumventing non-differentiability of discrete polyline approximations; (2) neural network parameterization of optical surfaces with bounded outputs ensuring physical feasibility; (3) CDF-matching objectives using stereographic coordinates for far-field intensity distribution matching. Currently extending these methods to handle finite extended sources and rotationally symmetric 3D systems, comparing neural approaches against deconvolution baselines.