CTH hydrocode, developed at Sandia National Laboratories, models the shock physics/dynamics of a material under an impact or other extreme environments. These simulations can be performed at the mesoscale through fine meshing and capture how local microstructure features affect the shock response. However, these fine meshes increase computation cost and limit the resolution and domain size that can be modeled, subsequently limiting the microstructure-to-property design capabilities. Machine learning solvers are used to significantly reduce the computation cost, but often comes with a reduction of accuracy, loss of high-frequency features, or limitations in modeling the temporal dynamics due to error propagation (as commonly observed in recurrent neural networks). In this work, several machine learning models are developed to act as a reduced-order model for simulating the shock initiation of a microstructure with a focus on maintaining the high-frequency features needed to capture different shock responses as a result of different microstructure characteristics. An autoencoder-type network is used to compress the physical fields, which are then passed to a different network that will learn the temporal dynamics in the compressed state. This presentation will focus on how to best maintain high-frequency features during the compression stage of training by means of tuning data processing steps, loss functions, model architecture, latent space manipulation, or post processing steps. Ensuring that high-frequency features are captured in the compression stage leads to better performance in the temporal dynamics model, yielding an overall more accurate model for simulating how the microstructure affects the shock response. SNL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525. SAND Number: SAND2026-17024A