arXiv:2611.09339v1 Announce Type: new Abstract: Accurate approximation of soliton solutions of the Korteweg--de Vries (KdV) equation remains challenging for physics-informed neural networks (PINNs) due to localized wave structures, high-order spatial derivatives, nonlinear interactions between solitons, and difficulties in long-time prediction. In this work, we address these challenges using a scaling PINN framework to approximate KdV soliton solutions. The framework transforms the governing equation into a normalized form, approximates its solution using a PINN solver, and reconstructs the solution in the original variables through inverse scaling. We introduce scaling transformations and PINN solvers tailored to one- and two-soliton problems. For the one-soliton case, we employ a PINN solver that incorporates a traveling-wave coordinate to capture soliton propagation. For the two-soliton case, we employ a two-step training procedure for the PINN solver. The first step learns the individual soliton profiles and estimates their propagation speeds, while the second step uses the estimated speeds to construct traveling-wave coordinates in the PINN solver and approximate the full two-soliton solution, including nonlinear interactions. We also introduce soliton-informed sampling (SIS), which places additional collocation points around the soliton trajectories. Numerical results show that the scaling PINN framework accurately captures one-soliton propagation, resolves two-soliton interactions, and preserves soliton profiles during long-time extrapolation in both cases. An additional three-soliton experiment suggests the applicability of the proposed two-step training procedure to multi-soliton solutions.