Precise optical alignment underlies systems ranging from fiber coupling and fluorescence microscopy to solid-state quantum platforms, yet it is still done largely by hand: researchers iteratively adjust mirrors and lenses based on signal feedback until performance is acceptable. This process is slow, non-quantitative, and dependent on individual expertise, which limits reproducibility. My project builds an autonomous alignment system that pairs programmable hardware control with intelligent optimization. The optical setup will be exposed through a uniform software interface so an AI agent can drive actuators and read detectors in real time-building on the same kind of autonomous control infrastructure used for photoluminescence excitation (PLE) measurements on diamond color centers-while a physics-informed surrogate model predicts measurable outputs from a given misalignment state. The surrogate is pretrained on optical simulations to encode the propagation physics, then calibrated online against real measurements; because it is differentiable, it supplies gradients that let an optimizer search the high-dimensional alignment landscape directly. I will first validate the system on a low-dimensional, near-convex case such as fiber coupling, then extend it to a harder, non-convex regime-sharpening the point spread function of single quantum emitters such as silicon-vacancy (SiV) centers, where the objective is a structured image rather than a scalar signal. The central question is whether a simulation-pretrained, experimentally corrected surrogate can guide alignment from poorly aligned states where classical methods fail.

What I like the most about my SuperUROP project is that the system only works if the physics and the machine learning are both right at once. Building a model that respects what we already know about optics while still learning what the simulation gets wrong is exactly the kind of challenge I find worth sinking a year into. A SuperUROP gives me the time and independence to do it properly-and I want to push it far enough to publish, and to come out with much sharper instincts for research.