Machine learning (ML)–based surrogate modeling offers a promising avenue for accelerating high-fidelity Computational Fluid Dynamics (CFD) simulations; however, even with recent advances in physics-informed learning, achieving physically meaningful predictions remains particularly challenging for complex fluid flow scenarios, such as compressible multiphase flows. In this work, we present a detailed benchmark study on a high-fidelity shock–droplet interaction dataset. Physics is incorporated through two complementary mechanisms. First, we inject non-dimensional flow parameters, such as the shock Mach number, by modulating layer normalization of the models via parameter-dependent affine transformations. These conditioning parameters encode governing properties of the system and guide the learning of trajectory-dependent dynamics. Second, we benchmark physics- and structure-aware training loss formulations, including Sobolev losses, interface-aware losses, SSIM-based perceptual losses, and wavelet-based spectral losses. Across a range of model architectures and parameter counts, both strategies consistently improve inference accuracy relative to purely data-driven baselines. Conditioning-based training further yields substantial data efficiency: through systematic ablations at 100%, 75%, 50%, and 25% dataset sizes, we show that models trained on only 25% of the data with parameter conditioning outperform unconditioned models trained on the full dataset. All benchmarks are conducted using an in-house, High Performance Computing (HPC)-ready ML training pipeline, Neptuna, designed for scalable distributed training and validated across multiple high-performance computing environments, including both NVIDIA- and Intel-based GPU architectures.