A physics-constrained surrogate for subcooled flow boiling in low-GWP immersion cooling microchannels: FiLM conditioning and gradient-balanced training

In liquid immersion cooling with the low-GWP refrigerant R1233zd(E), the onset of nucleate boiling (ONB) governs both the wall temperature and the thermal margin. A real-time digital twin needs a fast differentiable surrogate of the boiling closure rather than repeated evaluation of it. We present one: a physics-constrained network for a micro-finned R1233zd(E) channel, trained on correlation-generated data over 81 operating scenarios. The network conditions its coordinate backbone by feature-wise linear modulation and balances the physics and data losses from their measured gradient norms, replacing a hand-tuned constant; with fixed weights the energy residual absorbs 99.2% of the parameter gradient and the RMSE degrades by 67%. Across all 81 scenarios the median wall-temperature RMSE is 0.52 K against 2.37 K for a fully specified Sato–Matsumura baseline, ONB classification reaches an F 1 score of 99.4% against a 64.1% majority baseline, and the onset location is recovered to a mean absolute error of 0.15 mm, its presence identified correctly in every scenario. Leaving out an entire operating level, the constrained model is more accurate than an otherwise identical data-driven network in all four cases, by more than the seed spread in three. It is not uniformly more robust: beyond the trained pressure range it is 17 K worse, so input clamping is mandatory. Inference costs 0.64 ms on a CPU for a five-member ensemble. Against published R1233zd(E) measurements the ONB criterion reproduces the onset superheat to 1.30 K. The formulation has no dry-out model and over-predicts once dry-out begins.