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
physics.ao-ph · cs.AI · cs.LG
Quick verdict
Dainelli et al. distill Green AI guidance into an 11-question, stage-tagged ESM checklist plus metrics; a useful, well-sourced bridge, but unvalidated, thin at Deployment, and in need of copyediting.
Summary
This position/checklist paper from BSC's Earth Sciences department distills the Green AI literature into a practical checklist for ML/AI applications in Earth System Modeling (ESM). Eleven questions (G1-G11) are tagged to the earliest actionable stage among six (Scoping, Data, Training, Eval, Reporting, Deployment), with four starred as core (G3, G5, G8, G10); an Appendix self-assessment card (Done/Aware/No/N/A) is available as supplementary material. Each question is grounded in cited quantitative examples: AIFS produces a 10-day forecast in ~2.5 minutes on one GPU; Aurora fine-tuning outperformed from-scratch training by 54% on CAMS targets; a GraphCast adaptation took 37 GPU-days; Aurora reaches 37.2% of H100 peak versus 1.03% for GraphCast and 0.022% for SFNO; synthetic dataloaders raise SFNO/Pangu throughput 44x/10x; a Pangu rework cuts computation 20-30% and model size 40%. Section 3 catalogs eleven metrics (FLOPs, kWh, GPU-hours, TDP, utilization, CO2eq, grid intensity, PUE ~1.57, embodied carbon, PSF, energy-to-accuracy, storage footprint) in Table 1, mapped to stages, checklist items, and references. Main limitations: purely editorial — the checklist is never piloted on a real project; coverage is uneven (Deployment has one question); and editing artifacts (a duplicated G3 passage, citation-number slips, inconsistent author spellings) need fixing. All quantitative figures are literature-derived, not independently verified.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Green AI principles, dispersed across the ML literature, can be operationalised for Earth System Modeling as an 11-question checklist tagged to the earliest actionable stage of a six-stage project pipeline.
- Tracking and reporting must cover the whole R&D cycle, not the final run: final training accounted for only 9.6% of OpenAI's 2024 R&D compute, and of twelve surveyed frontier weather/climate models three report no compute figure at all.
- Reuse and fine-tuning of pretrained Earth-system foundation models are dramatically cheaper than from-scratch training (e.g., a GraphCast adaptation in 37 GPU-days; Aurora fine-tuning outperforming from-scratch by 54% on CAMS targets).
- Low-utilisation and I/O artefacts, not architecture, dominate much of the waste: reported sustained-throughput fractions range from 37.2% of H100 peak (Aurora) down to 0.022% (SFNO), and data-layout fixes projected a 54-hour run down to 1.3 hours on identical data.
- A dozen established metrics — from FLOPs and kWh to PUE, embodied carbon, PSF, and data storage footprint — suffice for ESM practitioners to estimate and report project-level energy and carbon impacts without computer-science specialist background.
Strengths
Questions are anchored to the earliest actionable stage, not generic principles
Sec. 2 assigns each of G1-G11 to one of six pipeline stages (Fig. 1), explicitly tagging the earliest stage since "the cost of revisiting an issue grows the further downstream it is caught"; four core questions (G3, G5, G8, G10) give practitioners a triage order, and the Appendix self-assessment card (Table 2, Done/Aware/No/N/A) makes it operable as a form.
Every question is grounded in cited, quantitative ESM-specific evidence
G1 cites AIFS (10-day forecast in ~2.5 min on one GPU vs. thousands of CPUs for IFS) and Samudra (~150x faster ocean emulation); G2 quantifies reuse (Aurora fine-tuning beats from-scratch by 54% on CAMS targets; Subich's GraphCast adaptation at 37 GPU-days); G8 turns profiling into numbers (37.2% of H100 peak for Aurora vs. 1.03% GraphCast vs. 0.022% SFNO; dataloader swap lifts SFNO/Pangu throughput 44x/10x; a 54h to 1.3h training-run saving from data-layout consolidation).
The experiment-tracking argument (G5) is unusually well-evidenced
The metrics section is practical and calibrated to what ESM researchers can actually obtain
Honest scoping of the tool's own limits
Weaknesses
No evaluation of the checklist itself
Editing defects that should not survive camera-ready
Uneven depth across stages the paper itself calls necessary
Quantitative anchors are literature-derived and presented without their conditions
Positioning claim about ESS practice is asserted, not evidenced
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a position/checklist paper from the Earth Sciences department of the Barcelona Supercomputing Center (arXiv 2609.00847, 1 Sep 2026). Its claim is narrow and honest: Green AI principles (Schwartz et al.'s Red AI / Green AI framing) are dispersed across the ML literature, and ESM practitioners need them packaged into a stage-tagged tool with domain-specific examples. The deliverable is (i) an 11-question checklist (G1-G11) mapped to six pipeline stages (Fig. 1), four starred as core (G3, G5, G8, G10); (ii) a metrics taxonomy (Sec. 3, Table 1) of 11 quantities from FLOPs to data storage footprint; (iii) a fillable self-assessment card (Table 2) as supplementary material. No new experiments are performed, and none are claimed — the paper's epistemic load sits entirely on how well it selects, contextualises, and attributes existing evidence.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the A Checklist for Green AI in Earth System Modeling paper about?
It converts scattered Green AI (energy- and carbon-aware machine learning) guidance into an 11-question checklist for researchers applying ML/AI within Earth system modeling — climate, weather, ocean, and related applications. Each question (G1-G11) is assigned to the earliest project stage where it can act: scoping, data, training, evaluation, reporting, or deployment, with four marked as core. It also lists standard metrics (FLOPs, kWh, CO2eq, PUE, embodied carbon, etc.) and tools (CodeCarbon, CarbonTracker, Green Algorithms, MLflow, W&B) for measuring and reporting a project's footprint.
How much energy do AI weather models use compared to traditional forecasting?
Drawing on Rieutord et al. (2026), the paper notes AI weather emulators have far cheaper inference than physics-based numerical prediction — e.g., ECMWF's AIFS produces a 10-day global forecast in about 2.5 minutes on a single GPU versus thousands of CPU-hours for the IFS — but their training carries significant, highly variable costs, so the net carbon benefit depends on how much the model is used and how honestly training-plus-R&D compute is accounted.
What tools can measure the carbon footprint of ML training?
The paper highlights free, open tools: CodeCarbon (Python-integrated emission estimates), CarbonTracker (tracks and predicts training footprint), the Green Algorithms calculator and GreenAlgorithms4HPC (broader hardware including HPC), and the Electricity Maps tool for real-time grid carbon intensity by region. It also recommends pairing energy measurements with the local grid carbon intensity and the data center's PUE (global average ~1.57) to convert kWh into CO2eq.
Is there a working checklist or self-assessment for sustainable AI research?
This paper provides one: eleven questions grouped into six project stages, four flagged as core (input feature necessity, experiment tracking, scalability profiling, full carbon reporting), plus a fillable self-assessment card (Done / Aware / No / N/A) available as supplementary Markdown. It is deliberately domain-tuned for Earth system science — covering, e.g., data granularity choices and physical-consistency validation — and is intended as a living tool that groups can adapt.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- G6 citation-number mismatch (Green Algorithms calculator tagged [29] rather than [31]) could partly stem from PDF text extraction; flagged as 'apparent' pending the published PDF.
- The duplicated Grundner et al. passage in G3 may be a typesetting or extraction duplication rather than an authoring error; either way it requires fixing before release.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.