Photonic microchips hold immense potential for the future of information society, yet waveguide crossings threaten to bottleneck their performance. To optimize the performance of waveguide crossings to minimize losses, inverse design techniques are often used. However, traditional algorithmic inverse design is highly computationally expensive, while often generating less robust geometries. This project proposes and executes a novel hybrid inverse design algorithm that integrates deep learning by using neural networks as a fast forward solver. Additionally, we propose a novel design methodology that generates elegant geometries which are highly robust against fabrication randomness. Our methodology can be applied to any waveguide crossing of crossing angle 90°, and we provide theoretical justifications in addition to empirical validation of it. We trained our predictive neural network model on 2800 data points. The model demonstrates highly accurate forward simulation with a MSE in the order of 10–6 to 10–7, while rapidly approximating simulations 5 orders of magnitude faster than exact solving of Maxwell’s equations. The final design generated by the inverse algorithm demonstrates a insertion loss of 1.64 dB, with an ultra-compact footprint of only 2.5 μm × 2.5 μm. Future tweaks to the design methodology as well as different settings can increase the subspace of searchable designs for further performance gains.