Neural surrogates that forecast a physical field from sparse, noisy sensors rely on two kinds of inputs: a spatial window of nearby sensors and a history of past snapshots. The size of both inputs can affect model performance, but the standard approach has been to select them empirically.
According to the arXiv abstract for paper 2610.12723, practitioners choose these input sizes by comparing fitted models with each other. The preprint appears to address this selection problem directly, though the abstract is cut off before stating any specific findings or recommendations.
Because the source only provides the framing, this digest cannot report the paper's conclusions. The significance lies in the question itself: how to systematically size inputs for partially observed fields rather than relying on ad hoc model comparisons.