A Checklist to assess the energy and carbon impacts of ML/AI applications in Earth System Modeling

arXiv:2609.00847 · physics.ao-ph, cs.AI, cs.LG · Submitted 2026-09-01 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "A Checklist to assess the energy and carbon impacts of ML/AI applications in Earth System Modeling".

Jane: The paper was written by Filippo Dainelli, Amirpasha Mozaffari, Marina Castaño, Gaya i Àvila, Lluís Palma Garcia et al. from Earth Department, Barcelona Supercomputing Center (BSC).

Tom: Stay tuned as we take you through the paper and discuss its implications.

Jane: We also have Lu with us today — senior AI researcher at Tsinghua.

Tom: We also have Meng with us today — lead engineer at a mysterious AI startup.

Jane: We also have Lalam with us today — the in-house Large Language Model.

Tom: Alright, let's get started.

Summary of the Paper: Jane: The authors summarize their approach by providing a detailed roadmap that tracks every single stage of developing an ML/AI application, from the initial idea all the way to sharing the final results.

Tom: This isn't just a list of steps; it’s a structured guide that follows six specific stages: Scoping, Data, Training, Evaluation, Reporting, and Deployment. It’s designed to make sure we see the whole picture.

Lu: It’s incredibly helpful because in real-world Earth system modeling projects like those used by ECMWF or the CAMS targets mentioned in the literature, things are rarely linear; they often require going back and forth between these stages to refine results.

Meng: This segmentation is a practical framework that forces us to consider costs across the the entire timeline, meaning resource consumption doesn't just start during training. The cost begins right when you decide on your initial scope of work.

Lalam: By visualizing the project this way, we gain a much clearer picture of the total environmental impact required to achieve any given result, which is something that was often completely invisible before this study.

Tom: The core message here is that by asking key questions at the very beginning—during Scoping—we can minimize the enormous waste of having to revisit and fix an issue later on when we’ve already spent massive amounts of time and compute power.

Jane: For example, if we define our model requirements poorly in the first phase, fixing that mistake after a multi-week training run is exponentially more difficult and wasteful than adjusting the initial assumptions right away.

Lu: It allows us to identify those critical trade-offs early—we can weigh model complexity against resource consumption right at the start, before committing to an expensive experiment that might prove overly ambitious for current hardware capabilities.

Meng: I see this as a tangible tool preventing wasted compute cycles by analyzing scalability and hardware utilization early on, helping determine if a model is even worth attempting on our available cluster resources.

Lalam: By summarizing the process this way, the paper is genuinely encouraging a culture of transparency around the total environmental cost associated with scientific research, which is absolutely critical for rebuilding trust in science itself.

Improvements Suggested by the Paper: Tom: We’ve established that "A Checklist to assess the energy and carbon impacts of ML/AI applications in Earth System Modeling" provides a fantastic structure. Now, let’s look at some of the actionable advice it suggests—the specific ways we can improve our approach.

Jane: The checklist offers incredibly practical advice, particularly about how to handle large datasets and how we can be much more careful with the sheer volume of data we feed into those powerful foundation models.

Lu: What struck me was the suggestion that instead of always aiming to train a model entirely from scratch, we should prioritize investigating whether there's already a large, pre-trained foundation model that can be used as a baseline, and then fine-tune it when it is technically feasible.

Meng: That’s a huge paradigm shift in thinking for many engineers who are accustomed to the "train everything" mentality. Reusing existing knowledge bases allows us to explore complex dynamics much faster and more efficiently than starting from zero every single time.

Lalam: I think this focus on fine-tuning is such an important step, allowing us to move toward a more efficient use of resources while still achieving highly accurate results in our climate simulations.

Tom: The checklist asks us to be very deliberate about feature selection too, advising us that we shouldn't just throw every possible data point into the model.

Jane: It encourages us to ask if we truly need all the features we are training on, or whether only those that carry the strongest signal are necessary for our specific task.

Lu: And it suggests running ablation studies and sensitivity analyses during evaluation—testing if a subset of features performs just as well as the using all twenty-four candidate inputs, for instance.

Meng: That’s a measurable way to reduce dimensionality; if we can achieve the same performance with fewer input variables, that directly translates into less computational time and resource usage.

Lalam: It gives us a clear path toward making our models lighter and more efficient without sacrificing the scientific rigor that makes them valuable.

Tom: And we’re not just cutting data; we are also encouraged to use tools like hyperparameter optimization instead of relying on trial and error, which is a massive time saver for researchers.

Jane: It points out that using automated tools can substantially reduce the number of exploratory training runs needed to find a good configuration, which saves energy and time.

Lu: The paper emphasizes that we need to be tracking our experiments with tools like MLflow or Weights and Biases from start to end, ensuring we don't redo work due to lost changes.

Meng: That tracking is essential for calculating the total R andD compute spending, because the final training run often accounts for only a tiny fraction of the total development footprint.

Lalam: It’s about creating a process that ensures efficiency is integrated into every single step, making sure our scientific curiosity doesn't result in an unsustainable environmental toll.

Paper discussion segment 3: Tom: So, we’ve looked at the structure and the improvements; now we are going to discuss what this paper means for the field—its broader implications for global climate science.

Jane: The checklist implies that sustainability can no longer be treated as an optional side project; it must become a core, measurable metric woven into every phase of model development.

Lu: What I find most striking is the idea that in our future research, the ideal climate scientist will need fluency not only in atmospheric chemistry but also in computational resource management and carbon accounting.

Meng: That’s a significant shift toward accountability; it forces us to think about how much energy was actually consumed by our models, rather than just assuming that computational overhead is an unavoidable cost of simply running things.

Lalam: By quantifying the carbon impact at every step—from data curation to deployment—it brings those hidden costs into sharp focus, which operationalizes what it means to be ethical in science.

Tom: And this has huge implications for how grant proposals are written; if sustainability becomes a mandatory criterion for funding bodies, it fundamentally changes the criteria for academic success.

Jane: It moves the conversation from "Can we build this model?" to "Should we build this model, given its environmental cost versus its scientific return?" That is a massive change in intellectual rigor.

Lu: The authors are pushing us toward a new breed of interdisciplinary researcher who understands these trade-offs better than ever before.

Meng: It also suggests that computational infrastructure needs to evolve; we can't just build bigger supercomputers, we must build smarter ones optimized for efficiency, not just brute force power.

Lalam: It’s about shifting the entire paradigm from "more compute equals more knowledge" to achieving the most efficient possible outcome for a reliable piece of knowledge.

Tom: Ultimately, this checklist provides the necessary framework and vocabulary to have these tough conversations with policymakers and funding agencies, giving us credible language to demand change.

Jane: It means that when we see any new methodology proposed in climate modeling literature, we should be asking: "What does this checklist say about its energy footprint?"

Conclusion: Tom: Looking back over everything we’ve covered today, it's clear that "A Checklist to assess the energy and carbon impacts of ML/AI applications in Earth System Modeling" is doing more than just a set of guidelines; it’s fundamentally changing how we think about computational science.

Jane: Exactly. It forces us to view the entire lifecycle of an AI model—from initial idea all the way to final deployment—through the lens of sustainability, making resource accountability central to our scientific process.

Lu: I appreciate that the framework is so comprehensive; it genuinely feels like a holistic approach that covers both the methodology and the ethical implications of deep learning in climate science.

Meng: What really sticks with me is how it elevates operational cost—the energy required to run these massive models—to the same level of importance as model accuracy itself.

Lalam: It’s such a valuable piece of guidance because it doesn't tell us *what* the answer should be, but rather gives us the necessary questions to ask ourselves before we start building anything complex.

Tom: And those questions are what truly make the difference between academic curiosity and responsible scientific advancement for our field.

Jane: It helps anchor the conversation in practicality, giving researchers something tangible to adopt right away, rather than just theoretical ideals of sustainability.

Lu: I think its greatest contribution is normalizing this discussion; making carbon impact a routine part of a project's initial planning stages across all major research institutions.

Meng: It signals that the community is mature enough to require and embrace this level of scrutiny before deploying powerful AI tools into global modeling efforts.

Lalam: This checklist ensures that groundbreaking science doesn't accidentally leave behind an unsustainable environmental footprint, which is vital for long-term trust in our scientific community.

Tom: Ultimately, this entire effort underscores the need for rigor across all stages, ensuring that sustainability is built in from the very beginning of every project.

Jane: To summarize its massive implications: this paper provides a much-needed blueprint for responsible AI development within global climate modeling initiatives.

Lu: It's an essential framework, making powerful science possible while keeping our environmental impact firmly in mind.

Meng: We should all keep the guidance from "A Checklist to assess the energy and carbon impacts of ML/AI applications in Earth System Modeling" close at hand as we move forward with our own work.

Lalam: Because it truly is a vital resource that guides us toward better, greener science for everyone.

Filippo Dainelli, Amirpasha Mozaffari, Marina Castaño, Gaya i Àvila, Lluís Palma Garcia, Alessio Melli, Oscar Dimdore-Miles, Amanda Duarte

Earth Department, Barcelona Supercomputing Center (BSC)

physics.ao-ph, cs.AI, cs.LG

Submitted: 2026-09-01

Updated: 2026-09-01

Code: https://github.com/keras-team/keras-tuner

Importance score: 95/100

The gist: The paper presents a comprehensive "Self-Assessment Card" designed as a practical companion to a checklist for assessing the energy and carbon impacts of Machine Learning (ML) and Artificial

Key concepts

The Six Stages
The paper provides a structured roadmap covering Scoping, Data, Training, Evaluation, Reporting, and Deployment. This framework forces researchers to track resource consumption across the entire project timeline. It ensures that environmental costs are considered even before the main training phase.
Fine-Tuning
This approach suggests avoiding training models entirely from scratch. Instead, researchers should use a large, pre-trained foundation model as a baseline and then fine-tune it for their specific task. This method is significantly more resource-efficient than starting from zero.
Feature Selection
This advice encourages researchers not to include every possible data point in a model. It suggests testing if a subset of features performs as well as using all inputs, often through ablation studies. This reduces dimensionality and saves computational time.
Sustainability in Science
The discussion emphasized that environmental impact cannot be an optional side project. Sustainability must become a core, measurable metric integrated into every phase of model development. This shifts the focus to considering the environmental cost versus the scientific return.

Terminology

Summary

The paper presents a comprehensive Self-Assessment Card designed as a practical companion to a checklist for assessing the energy and carbon impacts of Machine Learning (ML) and Artificial Intelligence (AI) applications within Earth System Modeling. Given the increasing reliance on computationally intensive AI models—such as those used in weather forecasting and climate prediction—this tool is crucial for ensuring that technological advancements are aligned with sustainability goals, thereby guiding researchers to develop green AI practices.

Scoping and Design Considerations

The initial phase of developing an Earth System Model utilizing ML requires rigorous upfront planning to minimize future computational waste. Key scoping questions guide the researcher toward establishing a clear justification for the model's environmental footprint. These include:

  1. Justifying the net carbon cost vs. benefit of the proposed AI application (G1).

  2. Determining if it is more efficient to fine-tune instead of training from scratch (G2).

  3. Investigating the necessary input features and data sources (G3).

  4. Ensuring that the data granularity is appropriately matched to the specific task needs, preventing over-specification or under-utilization of information (G4).

Data Management and Training Efficiency

The subsequent stages focus on optimizing resource use during model development. Diligence in this phase requires tracking all experimental iterations and understanding the underlying computational requirements. Researchers must address:

  • Tracking all experiments to maintain a clear record of resource consumption (G5).

  • Implementing methods to track carbon footprint with open libraries during training (G6).

  • Utilizing hyperparameter optimization tools to avoid exhaustive, wasteful searches (G7).

  • Conducting a thorough scalability and utilisation analysis to understand how model performance changes relative to computational resources (G8).

Model Selection, Evaluation, and Deployment

The final stages ensure that the developed model is not only accurate but also responsible in its deployment. The checklist mandates a shift toward efficiency during testing and sharing. Key practices include:

  • Testing simpler models and analyzing data subsets before scaling up to full-scale simulations (G9).

  • Generating a comprehensive Model/Data card with full carbon footprint reporting to ensure transparency for the scientific community (G10).

  • Commitment to responsible knowledge sharing by agreeing to share weights, code, and datasets alongside the findings (G11).

Overall, the framework provides a structured methodology for moving beyond mere performance metrics. It forces researchers to adopt a lifecycle perspective, treating energy consumption and carbon emissions as first-class variables alongside accuracy. By systematically addressing these eleven checkpoints across Scoping, Data handling, Training, Evaluation, and Deployment stages, the tool aims to foster sustainable ai practices in the critical field of climate science.

Improvements for AI systems

This analysis indicates that current AI systems applied to Earth System Modeling (ESM) lack mandatory energy accountability, often suffer from over-parameterization, and fail to integrate physical constraints efficiently.

Based on the synthesis of these references—particularly those concerning Green AI (Strubell et al., Wu et al.), model specialization (Zhao & Shen), and rigorous ML lifecycle management (The Self-Assessment Card)—I propose three major, integrated architectural improvements.

Improvement: The core architecture must be fundamentally redesigned to prioritize computational efficiency from the outset, moving beyond post-hoc energy audits. This involves integrating advanced model compression and quantization techniques directly into the training and inference pipeline.

Technical Mechanisms:

  1. Structured Pruning Integration: Implement dynamic structured pruning (as suggested by Zhao & Shen) during pre-training to eliminate redundant weights and connections that do not contribute significantly to the physical signal, dramatically reducing model size without substantial accuracy loss.

  2. Quantization and Low-Bit Inference: Mandate the use of low-bit quantization (e.g., INT8 or lower) for all deployed inference stages across tensor operations, minimizing memory footprint and accelerating computation on edge/distributed GPU clusters.

  3. Adaptive Resource Allocation: Integrate a real-time energy monitoring module that dynamically scales model complexity (e.g., reducing the number of transformer layers or attention heads) based on the required predictive fidelity vs. available computational budget, preventing unnecessary over-computation during routine runs.

What the Improved AI System Can Do:

  • Achieve Sustainable Forecasting: Produce high-resolution, subseasonal to decadal (S2S) climate forecasts with an energy consumption profile that is orders of magnitude lower than current state-of-the-art models.

  • Provide Energy Cost Estimates in Real Time: During any training or inference run, the system can output a precise estimate of the kWh consumed and the associated carbon cost (e.g., kg CO 2-eq), allowing researchers to make informed trade-offs between accuracy and sustainability.

Sources

Related papers