Configuring, Training, and Deploying AI Applications Through Integrated Gateways and Frameworks for Scientific Field Research

Abstract Adapting Artificial Intelligence (AI) workflows into scientific tools for research requires bridging the gap between general-purpose AI tools and the realities of field research. This paper presents a co-designed paradigm that distributes AI activities across a continuum of cyberinfrastructure-from resource-constrained edge devices to centralized HPC and cloud environments. At the edge, models are fine-tuned for local conditions and deployed for real-time tasks such as device resource optimization, detecting events of interest, and triggering event-based processing. At the center, large-scale training, evaluation, and version management ensure long-term accuracy and reliability. We outline the stages of AI adoption exposed to scientists—data preparation, model selection, training, evaluation, deployment, and monitoring—and show how middleware can simplify this lifecycle. Our work is guided by three research questions on: (1) transforming general AI workflows into scientific tools, (2) integrating edge devices into real-time decision-making, and (3) ensuring sustainable, versioned machine learning (ML) models in dynamic research contexts.