T1-2 - Workflows and Network Architectures
AUREIS Thrusts 1 and 2 focus on the software-to-architecture side of intelligent sensing: defining real-time workflows that can run within resource-limited sensing networks, and designing network and distributed-compute organization that maximize information extraction per unit of energy. The goal is dynamic, low-latency operation with feedback loops that enable real-time scientific insight and adaptive experiment control.
Modern detectors and sensing systems are often designed for worst-case signals and then stream raw data downstream, creating large mismatches between acquisition rate and useful information rate. In Thrusts 1–2 we invert that paradigm: start from the scientific decision-making needs, then co-design workflows and distributed architectures so computation happens where it is most efficient—at the sensor, within aggregation layers, and where appropriate in facility computing—while minimizing data movement and energy cost.
- Hardware/resource-aware ML/AI workflows for streaming data (reformulated for constrained edge resources)
- Distributed workflow partitioning across front-end ASICs, FPGAs/aggregation layers, and facility compute
- Dynamic data-flow design that adapts to changing experimental goals and conditions
- Network topology + compute organization trade studies to identify when to process locally vs aggregate vs centralize
- Energy + performance benchmarking methodology for heterogeneous hardware/software workflows
Interfaces: T1–2 provides requirements and targets (latency, bandwidth, compute/memory budgets, data-reduction goals) that drive T3–4 architecture choices and define representative data characteristics relevant to T5 sensor constraints.
Outputs: workflow definitions, benchmark methodology, and representative demonstrations.
