Evolutionary Optimization of Neural Architectures for Holographic Phase Prediction in Optical Tweezer Arrays

We present an evolutionary optimization framework for designing neural architectures for holographic phase prediction in optical tweezer arrays. The framework combines Fourier-transform-based network structures with genetic optimization of layer composition and connectivity, enabling optimization of prediction quality under an inference-time constraint. The optimized architecture achieves phase prediction quality comparable to several iterations of the Gerchberg–Saxton (GS) algorithm while significantly reducing the inference time. The resulting model generalizes to larger array sizes beyond the training set and can be directly implemented on a spatial light modulator. We present this as a proof-of-concept demonstration of the architecture-search methodology and of the physical realizability of the predicted holograms; a fully benchmarked model with large-scale statistical evaluation and real experimental validation is left for future work. Defect-free single-atom array assembly for neutral-atom quantum computation and quantum simulation provides the broader motivation for this line of work.