CA-FMLOps: Evaluating Edge Filtering and Carbon-Aware Scheduling in Sustainable MLOps
Abstract
Artificial intelligence can support climate mitigation, but the computing systems used to develop and operate it can also intensify energy demand, data movement, and greenhouse-gas emissions. This paper presents and evaluates CA-FMLOps—Carbon-Aware Frugal Machine Learning Operations—as a design-science framework for climate-technology systems that must operate under constrained energy, connectivity, compute, and financial conditions. The framework combines four controls: edge-first inference, event-triggered data transmission, carbon-aware scheduling of deferrable cloud workloads, and lifecycle measurement using SCI-aligned accounting. A 50-run Monte Carlo benchmark with 2,000 observations per run reports a 98.352% simulated operational-carbon reduction and a 99.985% simulated deadline-reliability rate for the full hybrid arm relative to the cloud-first baseline, while also exposing a 33.332% recall limitation. Keywords: sustainable AI, Green AI, MLOps, edge computing, carbon-aware computing, wildfire detection
