Scientists are increasingly incorporating machine learning (ML) and artificial intelligence (AI) techniques into their applications to accelerate and enhance scientific research and discovery across a wide range of disciplines. For example, machine learning has been successfully integrated into tools for weather forecasting, earth sciences, astronomy, high-resolution imaging, genomics, and molecular biology. However, the ever-growing size of scientific datasets results in prohibitive hardware resource costs, significantly complicating the deployment of these applications on high-performance computing platforms at scale. Lossy compression — a data reduction technique that significantly reduces dataset size by removing redundant or less important information — has proven effective for many scientific datasets, including those from cosmology and structural biology. Despite its promise, integrating lossy compression into AI-driven scientific applications remains a non-trivial challenge, requiring broad expertise in data compression and machine learning, as well as a deep understanding of application requirements, system considerations, and their interactions. These complexities hinder the adoption of this powerful data reduction technique in scientific applications. The overarching goal of this project is to address this gap by providing a cyberinfrastructure that seamlessly and adaptively integrates lossy compression into deep learning pipelines within scientific applications