Anyone who has spent time optimizing a fiber laser cavity, a Raman amplifier pump configuration, or a grating coupler knows the frustration: you run a simulation, adjust a parameter, run it again, repeat. For complex devices, a single FDTD or FEM simulation can take hours, and you may need hundreds of iterations before converging on something good.
ML-driven inverse design turns this around - you specify the performance you want and a trained neural network proposes the device structure directly, typically in milliseconds. We just published a comprehensive review (full open access: https://doi.org/10.1117/1.APN.5.1.014002) of how this approach is being applied across the full optical component stack relevant to communication and sensing systems: semiconductor and fiber lasers, Raman and semiconductor optical amplifiers, power splitters, gratings, fiber Bragg gratings, mode-selective couplers, optical fibers (FMFs, PCFs, hollow-core), and metamaterial absorbers.
Across 65 papers from 2019 to 2025, the results are genuinely impressive for individual components - NNs predicting Raman pump configurations with errors below 0.5 dB, laser parameter extraction running over 14,000× faster than traditional methods, and mode-selective coupler design cut from 4 hours to 45 seconds. But the field has a significant gap that I’d be curious to hear this community’s thoughts on: almost no one is optimizing multiple interdependent components together. Every paper treats devices in isolation. For real systems - especially anything involving integrated sensing and communication - you need your source, amplifier, splitter, and fiber to be co-designed, not independently optimized and then assembled. The cost of putting a full multi-component simulation pipeline inside a training loop is still too high, and no one has cleanly solved it yet.
Has anyone here worked on system-level photonic design optimization, or run into this bottleneck in practice? Would be very interested in the engineering perspective from people building real systems 
Full open-access paper: https://doi.org/10.1117/1.APN.5.1.014002 (Advanced Photonics Nexus, Jan/Feb 2026)