Source-linked AI summary
A Checklist to assess the energy and carbon impacts of ML/AI applications in Earth System Modeling
Filippo Dainelli, Amirpasha Mozaffari, Marina Castaño, Aina Gaya i Àvila, Lluís Palma Garcia, Alessio Melli, Oscar Dimdore Miles, Amanda Duarte
TL;DR
AI/ML is becoming increasingly computationally intensive in Earth System Modeling, while sustainable-practice guidance remains dispersed. The paper distills Green AI literature into a stage-based checklist, supported by metrics and software for assessing energy use and carbon footprint. Its contribution is to make existing principles actionable at the points in model development where decisions become consequential.
Problem
Earth System Modeling increasingly uses computationally intensive AI/ML, while practical guidance for reducing its environmental footprint remains dispersed across the literature.
Method
The paper develops a practical checklist organized around AI/ML project phases and complements it with environmental-impact metrics and open-source tracking software.
Results
The paper distills existing Green AI discussions into actionable questions, examples, metrics, and software guidance tailored to Earth System Modeling.
Takeaways & Limitations
Practitioners can assess and reduce AI/ML environmental impacts by considering relevant questions at appropriate development stages and tracking project energy and carbon costs.
Abstract
from arXiv · showhide
As machine learning and artificial intelligence find their way into nearly every aspect of climate, weather, and Earth system modeling, it is worth pausing to consider what our design decisions imply for the science and for the computational resources we consume. A growing body of literature addresses the ethical and sustainable development of ML/AI, yet translating these principles into day-to-day research practice remains a challenge as most of best practices are dispersed across multiple studies and commentaries. Here, we distill these discussions into a practical checklist that ML/AI and Earth system science practitioners can use to assess and reduce the environmental footprint of their own applications, organised around the successive stages of the model development pipeline. We complement the checklist with a selection of metrics drawn from the literature for estimating the energy consumption and carbon footprint of a project. For each question, we point to concrete examples and actionable suggestions from recent literature, aiming to bridge the gap between aspirational principles and the decisions researchers face at every stage of the development cycle.
1 Introduction
The paper argues that increasingly computationally intensive AI/ML models create environmental concerns in Earth System Modeling. It proposes a practical checklist and metrics to make Green AI principles actionable throughout project development.
- Increasingly large and computationally intensive models have improved capability while increasing architecture complexity, parameter counts, and resource use.
- Green AI evaluates computational efficiency alongside accuracy to reduce environmental footprint and lower participation barriers.
- The checklist applies Green AI principles specifically to Earth System Modeling, where AI models both study environmental systems and consume relevant computational resources.
- The proposed checklist is organized by AI/ML project phases and is complemented by metrics and open-source software for tracking energy consumption and carbon footprint.
- The paper presents environmental-impact metrics in a separate section for assessing AI/ML projects.
2 Checklist
The checklist assigns actionable questions to successive project stages, emphasizing early decisions about scope, data, training, reuse, efficiency, tracking, and reporting. Examples show that reuse, profiling, and complete experiment accounting can substantially change computational and environmental costs.
- Checklist design: Each checklist question is assigned to the earliest project stage at which it becomes actionable, because later correction costs more.
- Checklist design: The six stages follow an AI/ML project’s progression, with scoping setting fundamental trade-offs between model complexity and resource consumption.
- Scoping: Fine-tuning Aurora outperformed the same architecture trained from scratch by 54% on average across CAMS targets.
- Data and features: Feature selection can reduce input dimensionality, computational time, and resource use while preserving performance and improving transparency.
- Training and reporting: Experiment tracking links runs to code versions, prevents redundant runs, and estimates cumulative R&D compute rather than only final training cost.Final training represented 9.6% of OpenAI’s 2024 R&D compute spending, while two smaller developers reported 12.3% and 22.6%.
- Training and reporting: The same experiment can produce up to 30 times more CO2 emissions depending on the electricity grid, making location a consequential model-independent choice.
- Training and profiling: Profiling processor, memory, and I/O behavior reveals inefficiencies that wall-clock time alone misses.
3 Metrics
The paper presents metrics for assessing energy and carbon impacts across AI/ML project stages, from architecture choice and operational energy use to project-level and storage footprints.
- Scoping and design: FLOPs compare candidate architectures before training but cannot directly estimate energy or carbon because they are hardware agnostic.They identify which architectures are more computationally demanding during scoping and design.
- Operational metrics: Energy consumption in kWh measures electricity used by CPUs or GPUs during training, evaluation, and inference.Direct measurement can use tools such as NVML or CodeCarbon.
- Operational metrics: CPU-hours and GPU-hours provide coarse run-level measures that can be converted to kWh using manufacturers’ Thermal Design Power.TDP-based estimates may overestimate energy at low utilisation or underestimate it because they account only for accelerator power.
- Carbon accounting: CO2eq is estimated by multiplying energy consumption by grid carbon intensity, while PUE accounts for total data-centre energy relative to IT equipment energy.Both factors depend on computing location; the reported global average PUE is ∼1.57, though facility-specific values can differ.
- Project-level metrics: The Pragmatic Scaling Factor estimates total project footprint from a final run, while energy-to-accuracy compares energy or carbon cost with model performance.These project-level metrics can include failed experiments, debugging, hyperparameter searches, and ablations; storage footprint additionally covers energy for data retained on disk, including replication and backups.
4 Conclusion and Outlook
The paper turns Green AI principles into a stage-based checklist and complementary metrics for Earth system science, emphasizing early consideration of environmental costs. It presents the checklist as a non-exhaustive, adaptable tool for evolving practice.
- Conclusion and Outlook: The checklist makes existing Green AI principles actionable by attaching questions to the project stage where they are most consequential.The paper argues that early action is cheaper and that ESM choices such as resolution, domain, temporal coverage, and ensemble size amplify compute and carbon.
- Conclusion and Outlook: The checklist is non-exhaustive and intended as a living tool that research groups can adapt and the community can extend.The present version focuses on environmental sustainability, with future iterations potentially addressing equity, fairness, and trustworthiness.
A Self-Assessment Card
The self-assessment card provides a practical way to record responses to checklist questions, using four statuses for each row.
- A Self-Assessment Card: Each checklist question is marked Done, Aware, No, or N/A to support project self-assessment.A standalone fillable Markdown version is available as supplementary material.
- A Self-Assessment Card: The card identifies core questions with a star and provides one response field per row.This format is presented as a practical companion to the checklist.