Learning accurate models of dynamical systems typically requires large quantities of labeled data, which are often costly or impractical to acquire in real-world scientific and engineering applications. Although physics-based simulators offer abundant supervised data, transferring learned representations from simulation to real-world settings remains a significant challenge, particularly when observations in the target domain are only partially labeled.
In this work, we investigate transfer learning for time-series modeling under asymmetric supervision, a setting in which simulated trajectories are fully labeled while only sparse temporal observations are available in the real domain. To address this challenge, we introduce INR-FiLM, a FiLM-modulated Implicit Neural Representation architecture that combines a Fourier-encoded multilayer perceptron with a GRU-based modulation network. Preliminary empirical results demonstrate that INR-FiLM significantly reduces trajectory prediction error, even when real-world supervision is extremely limited.