Boundary-Condition Compliance in Learned Fluid Simulators: Boundary Particles over Distance Features and the Primacy of Geometric Fidelity

Learning-based simulators reduce the cost of particle-based fluid simulation, but autoregressive rollouts accumulate error over hundreds of steps. The failure that matters most is the loss of boundary-condition compliance: once fluid particles leave the domain, the remaining prediction is physically meaningless. This is usually treated as a training problem. We show instead that at a fixed model and a fixed training budget it turns on where the boundary is placed. We represent container walls as static boundary particles of the same type as rigid bodies, where the type acts as a multiplicative weight in message aggregation combined with a learned attention gate. This ports the boundary particles of smoothed particle hydrodynamics to a learned simulator without architectural change, preserving translation invariance. On Water-3D under a matched training budget, it reduces the 400-step domain escape rate from 35.8 ± 8.3 percent for a graph network simulator with clipped wall-distance features to 1.29 ± 0.32 percent. Moving the wall particles onto the uncorrected metadata box, 1 percent of the domain further in, collapses the rollout to 74.0 ± 9.4 percent escape and injects 2.8 times more energy than omitting walls entirely. Single-step validation loss does not track any of this: across three seeds it differs by 7.4 percent between two configurations whose escape rates differ 11-fold, and on the seed where the two losses are within 0.3 percent the escape rates still differ sevenfold. Boundary representation is therefore a design variable rather than an implementation detail, and we report error, conservation drift, and boundary violations jointly across the rollout horizon.