The context: I’m planning to deploy a novel hydrogen production system in 2026 and instrument it extensively to test whether hard-constrained PINNs can optimize complex, nonlinear industrial processes in closed-loop control. The target is sub-millisecond (<1 ms) inference latency using FPGA-SoC–based edge deployment, with the cloud used only for training and model distillation.
I’m specifically trying to understand:
Are there practical ways to enforce hard physical constraints in PINNs beyond soft penalties (e.g., constrained parameterizations, implicit layers, projection methods)?
Is FPGA-SoC inference realistic for deterministic, safety-critical control at sub-millisecond latencies?
Do physics-informed approaches meaningfully improve data efficiency and stability compared to black-box ML in real industrial settings?
Have people seen these methods generalize across domains (steel, cement, chemicals), or are they inherently system-specific?
I’d love to hear from people working on PINNs, constrained optimization, FPGA/edge AI, industrial control systems, or safety-critical ML.
I’m not hiring at this stage — this is purely to learn from the community and potentially collaborate on research or publications as data from the industrial pilot becomes available. I’m also happy to share findings as the project progresses.
If you have experience, references, or strong opinions here, I’d really appreciate your thoughts.