Neural surrogates for crystal growth dynamics with variable supersaturation: explicit vs. implicit conditioning

The deeper question this paper is really asking is whether a neural network can reliably infer hidden physical parameters from observation alone, or whether it needs those parameters handed to it explicitly. That distinction has consequences far beyond crystal growth: it maps directly onto the general problem of deploying physics surrogates in real industrial settings where sensor coverage is incomplete.

Recent Modelwire coverage has focused heavily on generalization failures in learned models. The 'Generalization in LLM Problem Solving' piece from April 16 showed that models can transfer spatially but collapse when problem complexity scales, and a similar failure mode is plausible here: implicit conditioning may work within the training distribution of supersaturation values but degrade at the edges. The nonlinear separation principle paper from the same date is also loosely adjacent, in that both works are probing structural conditions under which recurrent neural architectures remain well-behaved. Outside the archive, this work belongs to a growing body of physics-informed machine learning research aimed at replacing expensive numerical solvers in materials science.

If the implicit conditioning model maintains accuracy within five percent of the explicit model across supersaturation values outside the training range in follow-on ablations, that would be a meaningful signal that sequence-based inference is a viable path for deployable surrogates. If the gap widens significantly out of distribution, the explicit approach will likely dominate practical adoption.