A Zero-Shot Multi-Scope Life Cycle Assessment Framework for Machine Learning: Unifying Carbon, Water, and Network Footprints

Abstract

Machine learning workloads now consume a significant and increasing percentage of global data centers energy. To overcome this problem, researchers and developers of new AI applications and pipelines are forced to use post hoc techniques and approaches to power telemetry to measure the actual power consumption during model executions. However, these approaches present a crucial and structural problem. Profiling a model requires running it first, which means we waste energy trying to measure its models’ energy’s draw. Furthermore, these tools and approaches are affected by ambient cloud virtualization noise rather than the actual algorithmic efficiencies. We propose a zero-shot, multiscope Life Cycle Analysis (LCA) framework that bypasses the actual physical execution of a model by recursively traversing a Directed Acyclic Graphs (DAGs) and extracting topological depth of the tree-based ensembles. The engine maps the graph-level arithmetic intensity to physical hardware saturation ceiling via dynamically rooflining modeling and profiling. Our proposed framework analytically bonds operational carbon, scope 3 embodied manufacturing which is type of carbon emitted during the different manufacturing process of hardware, carbon resulting from network transmission and its associated penalties, and finally direct/ indirect water footprints. The evaluation study was carried out on diverse set of modularity across vision, tabular and time series benchmarks on diver set of CPU and CPU accelerators, the static proposed approaches achieves a profiling latency of 0.05 ms to 0.08 ms, this represents a massive reduction compared to the 510.83S required for full physical execution of models. Validation against hardware telemetry confirms a strong rank-orders preservation (spearman’s p = 0.8333 to 0.9286) across modalities. A multi-scope projection demonstrates that wireless transmission footprint can in some cases exceed the edge-inference carbon emission by over six orders of magnitude. This framework allows researchers to perform sustainable architectural exploration before committing physical compute resources, eliminating hidden carbon taxes of green AI model profiling.