GUINEAGRAD: A Differentiable Beam-Beam Interaction Simulator for Gradient-Based Collider Optimization

Making physical simulators differentiable unlocks gradient-based optimization. We present GuineaGrad, a JAX reimplementation of the Guinea-Pig beam-beam interaction code that computes automatic derivatives of luminosity and beamstrahlung observables via reverse-mode automatic differentiation. Validated against Guinea-Pig to within 2\% across three collider configurations, GuineaGrad enables gradient-based multi-objective optimization in a 6-dimensional parameter space where Nelder-Mead collapses entirely, producing only degenerate or divergent solutions. Gradient-based optimization becomes wall-clock faster than derivative-free alternatives at ${\sim}$ 6-10 parameters, establishing differentiable simulation as a practical tool for high-dimensional collider design.