A neural network that hard-codes hydrogen transport physics into its structure can predict dangerous gas crossover far more reliably than data-only models — even at pressures well beyond its training data.
Researchers built a physics-residual network, called PR-Net, for predicting hydrogen crossover in polymer electrolyte membrane water electrolyzers — the industrial systems that split water into hydrogen and oxygen under high pressure. Instead of asking a neural network to learn all the physics from scratch, PR-Net embeds three established laws (Henry's, Fick's, and Faraday's) as a fixed backbone, then trains a small residual layer to capture only the nonlinear effects those laws miss. Tested against 184 observations drawn from eight peer-reviewed studies across six membrane types, it reached an R² of 99.57% — with nine times lower prediction variability than either a standard neural network or a soft-constraint physics-informed neural network.
The extrapolation result is where things get interesting for industrial deployment. At 200 bar — two and a half times the maximum pressure in the training set — PR-Net held an R² of 94%, while the physics-informed neural network dropped to 68% and the plain neural network fell to 58%. That gap matters because real-world electrolyzer operators need to push systems beyond tested conditions, and hydrogen crossover is a direct safety hazard: too much gas migrating through the membrane can cause fires or explosions. The model also runs in about one millisecond on low-power embedded hardware, making real-time monitoring feasible without a server rack.
The hybrid approach — lock in what physics already knows, learn only what it doesn't — is a sensible counter to the current trend of throwing more data and larger models at every engineering problem. Green-hydrogen production is still looking for its cost curve; tools that reduce the experimental burden of safety validation could matter more than they look.