Explainable AI for Geoscientific Regression: Evaluating Sensitivity and Attribution Methods with the Lorenz-63 System and Applications

arXiv:2608.07406 · physics.ao-ph · Submitted 2026-08-07 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.

Kai: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: We're looking at the paper: "Explainable AI for Geoscientific Regression".

Mira: Here's the gist of the paper. Explainable artificial intelligence (XAI) methods are being evaluated for regression problems in geoscientific applications using an emulator of the Lorenz-63 system to understand their limitations and failure modes <ref:2608.07406#pg20>.

Kai: This work reviews XAI approaches through a structured framework, applying them to a chaotic model to ground the discussion before surveying their use across Earth system sciences <ref:2608.07406#pg4>.

Mira: Pruning redundant reasoning tokens The paper establishes that most existing XAI methods were developed for classification tasks and may not transfer well to the regression problems that dominate geoscientific applications <ref:2608.07406#pg4>.

Kai: The authors ground their discussion by applying a selection of methods to a machine learning emulator of the Lorenz 1963 system, an archetypal chaotic model that provides a tractable, physically meaningful setting for exposing the limitations and failure modes of general-purpose XAI in regression contexts <ref:2608.07406#pg4>.

Paper summary: Kai: So, looking back at "Explainable AI for Geoscientific Regression: Evaluating Sensitivity and Attribution Methods with the Lorenz-sixty-three System and Applications", the core message is that we need to be careful about assuming that XAI methods built for classification will just work when applied to regression problems in geosciences <ref:2608.07406#pg1>.

Mira: They argue it’s important to review these approaches through a structured framework, looking at both their theory and how they actually perform practically, especially since the Lorenz sixty-three system provides a good testbed for this > <ref:2608.07406#pg1>

Lev: The implication is that we need to be much more intentional about how we select our baseline states because you can't just pick an arbitrary starting point without considering its physical meaning >

Kai: And they give some practical advice, like don't assume methods transfer directly and always choose your reference states with intent, grounding them in something physically justifiable >

Mira: They also stress matching the explanation to stakeholder questions rather than trying to explain the whole model at once; tailor the explanation to what that specific user needs >

Lev: And they point out a big challenge: scaling these methods up to massive systems like weather models means we really need XAI tools that can handle huge numbers of inputs and outputs reliably >

Kai: The paper suggests that for operational use in high-stakes environments, we need to refine these methods and develop new techniques specifically suited for the unique characteristics of Earth system modeling >

Conclusion: Kai: So, we've looked at how different explainable AI methods work when applied to predicting continuous values in chaotic systems, specifically using an emulator of the Lorenz sixty-three system.

Mira: And what this paper is doing is taking a structured look at those methods—how they are built and what they actually do—before applying them to the bigger picture in Earth system science.

Lev: The authors are showing that most current XAI tools were made for things like classification, which isn't quite right for regression problems, where you're predicting a continuous number.

Kai: They use the Lorenz sixty-three system because it’s a chaotic model, and that gives them a clear place to test where these general AI explanations break down.

Mira: They examine several techniques like gradients and SHAP, and they find that the input variable labeled y has the biggest impact right near points where things diverge in the model.

Lev: But they also point out how important it is to be careful about what you consider a starting point or a reference state because that choice really matters for the explanation's reliability.

Kai: So, when you're building these models for things like weather or climate, the paper suggests that explanations shouldn't try to explain everything at once.

Mira: They want explanations tailored to what each stakeholder actually needs, not just one big answer trying to cover all the model complexity.

Lev: The big practical points they give are about picking your reference states with intention and making sure you only look at inputs that make physical sense for the system you're modeling.

Kai: It seems like the main challenge right now is figuring out how to make these explanations more reliable and scalable for really massive models, not just little test systems.

Mira: And they suggest that we need to develop new, domain-specific methods because the current routine tools have technical limitations when you move into high-stakes operational use.

Lev: That leads us right into thinking about how we can adapt these sensitivity methods—like adjoint methods—to fit these specific needs in the real world.

Department of Computer Science, University of Reading · Department of Meteorology, University of Reading · National Centre for Atmospheric Science (NCAS) · Met Office

physics.ao-ph

Submitted: 2026-08-07

Updated: 2026-11-04

Comments: 66 pages, 15+10 figures in main+appendix. This Work has been submitted to Artificial Intelligence for the Earth Systems. Copyright in this Work may be transferred without further notice

Code: https://github.com/ichiggs/Evaluating-XAI-Methods-L63

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 77/100

The gist: Explainable artificial intelligence (XAI) methods are being evaluated for regression problems in geoscientific applications using an emulator of the Lorenz-63 system to understand their limitations…

Key concepts

Regression vs. Classification
Regression predicts continuous values, unlike classification which predicts discrete categories. Most existing XAI tools are built for classification tasks, meaning they may not accurately show feature importance or influence when predicting continuous physical quantities in Earth science models.
Lorenz-63 System Emulator
This is a simplified computer model of the Lorenz system, a classic chaotic system used to simulate atmospheric dynamics. It serves as a tractable, physically meaningful environment to test the limitations and failure modes of general XAI methods when applied to complex regression problems.
Post-hoc XAI Methods
These are techniques applied *after* a model is trained to explain its predictions. Examples include Saliency and SHAP, which quantify how much each input variable contributed to a specific output prediction, helping users understand the model's decisions.

Terminology

Summary

Explainable artificial intelligence (XAI) methods are being evaluated for regression problems in geoscientific applications using an emulator of the Lorenz-63 system to understand their limitations and failure modes <ref:2608. p. 207406#pg20>. This work reviews XAI approaches through a structured framework, applying them to a chaotic model to ground the discussion before surveying their use across Earth system sciences <ref:2608. p. 107406#pg4>.

Pruning redundant reasoning tokens

The paper establishes that most existing XAI methods were developed for classification tasks and may not transfer well to the regression problems that dominate geoscientific applications <ref:2608.07406#pg4>. The authors ground their discussion by applying a selection of methods to a machine learning emulator of the Lorenz 1963 system, an archetypal chaotic model that provides a tractable, physically meaningful setting for exposing the limitations and failure modes of general-purpose XAI in regression contexts <ref:2608. p. 107406#pg4>. They organize XAI approaches into a structured framework examining both their theoretical foundations and practical behavior <ref:2608. p. 107406#pg4>.

How the training signal is built

The study uses a neural network trained to forecast continuous state variables in the L63 system, which is implemented with PyTorch and achieves a test error of 100× RMSE/σ ≈ 0.62% <ref:2608. p. 007406#pg40>. The training data is generated over a long trajectory (2,000 time units) and split into training, validation, and test sets at a 64:16:20 ratio <ref:2608. p. 807406#pg40>. The Adam optimiser with a learning rate of 0. p. 9001 and mean square error loss was used for training <ref:2608. p. 907406#pg40>.

Where the gains come from

The authors detail several post-hoc XAI methods applied to the L63 system, including gradient-based methods like Saliency, Input×Gradient, Integrated Gradients, SHAP, and Layer-wise Relevance Propagation (LRP) <ref:2608.07406#pg40>. These methods provide local sensitivities, attributions or relevance scores for individual predictions by quantifying each input variable’s influence on each output <ref:2608.07406#pg40>. The analysis shows that the input y-variable exhibits the highest importance near divergence points, followed by x, while z shows comparatively little importance <ref:2608. p. 807406#pg40>.

Linking XAI to model development and stakeholder needs

The paper frames the interaction between stakeholder needs and explanations within a simple model lifecycle framework oriented around the development and operational deployment of Earth system models <ref:2608.07406#pg40>. Different stakeholders engage with AI models in distinct ways at each stage of this cycle, and what constitutes a satisfactory explanation varies with the needs and interests of individual stakeholder groups <ref:2608. p. 107406#pg40>. The authors conclude that explanations should be tailored to specific stakeholder questions rather than attempting to explain a model in its entirety <ref:2608.07406#pg40>.

Practical recommendations

The key practical recommendations include:

  1. Do not assume methods designed for classification transfer directly to regression, as regression lacks a decision boundary and probabilistic output that naturally encodes uncertainty <ref:2608. p. 407406#pg40>.

  2. Choose your baseline or reference states with intentionality, grounding them in a physically meaningful, justifiable reference <ref:2608. p. 907406#pg40>.

  3. Prioritise inputs “familiar” to the model, as attributions from physically implausible inputs may lose their connection to realistic behavior <ref:2608. p. 107406#pg40>.

  4. Justify aggregation of explanation metrics or examine them in more detail, as removing sign information can obscure a feature’s sign of influence <ref:2608. p. 407406#pg40>.

  5. Match explanations to stakeholder questions, defining the question first and then selecting or designing an appropriate method <ref:2608.07406#pg40>.

Key challenges and future directions

The key challenges identified are:

  1. Refinement of existing XAI methods is needed, particularly regarding baseline selection and accounting for correlated inputs <ref:2608.07406#pg40>.

  2. Scaling to large systems like NWP requires XAI methods that can operate reliably on models with very large numbers of input and output features <ref:2608. p. 607406#pg40>.

  3. Stakeholder-specific usefulness requires adapting XAI methods to Earth system modelling’s specific characteristics and translating complex attribution maps into concise, meaningful summaries for operational use <ref:2608. p. 107406#pg40>.

Conclusion

The work concludes that technical limitations are often present when routine XAI methods are used in regression settings for Earth system applications, highlighting the need for continued development of domain-specific explainability methods suited to high-stakes operational contexts <ref:2608.07406#pg40>. The authors suggest that sensitivity-based approaches have clear analogues in well-established adjoint methods, and early bespoke techniques show considerable promise for interpreting AI models’ internal mechanisms <ref:2608.07406#pg40>.

Acknowledgments

The authors gratefully acknowledge the collaborative support and funding provided by the University of Reading and the UK Met Office, which enabled the work reported in this publication via the Met Office Academic Partnership framework <ref:2608.07406#pg40>. The code used for L63 illustrations can be found at https://github. p. 4com/ichiggs/Evaluating-XAI-Methods-L63 <ref:2608. p. 207406#pg40>.

Appendix A

Appendix A provides additional detail on illustrative L63 explanations, showing the structure of the full attribution and complete explanation structures for each method such as Saliency, Input×Gradient, Integrated Gradients, SHAP, LRP-ε rule, LRP-α1β0 rule, and LRP-γ rule <ref:2608.07406#pg40>. The structure of explanations is represented by a matrix of size M × N for any ML model with M inputs and N outputs <ref:2608. p. 407406#pg40>. The summed result of Eq. (3) is the column-wise summation of the absolute values in Fig. p. 8 A1, which serves as a summary measure of feature importance <ref:2608. p. 007406#pg40>. Individual input attribution vectors correspond to a row in Fig. p. 6 A1, such as if the output target is set to be zt+n then the vector returned consists of elements [g,h,i] <ref:2608.07406#pg40>. The complete explanation structures show how relevance scores sum to the model output through propagation rules like LRP-ε rule and LRP-γ rule <ref:2608. p. 207406#pg40>.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Incorporate an XAI method tailored for regression that explicitly accounts for correlated inputs, moving beyond assumptions of feature independence to avoid physically misleading attributions. This allows the system to provide more reliable diagnostic information when dealing with the rich correlation and covariance structures typical of geoscientific data.

  2. Implement a baseline selection mechanism that is physically meaningful and justifiable, such as a dynamical state of interest or a quantity relevant to the stakeholder’s question, instead of an arbitrary baseline which produces attributions that are difficult to interpret and potentially misleading.

  3. Design the system to prioritize inputs familiar to the model by using explanations derived from training data or ensemble members when off-manifold inputs are used, ensuring that explanations speak to the model's limitations rather than the phenomenon of interest.

  4. Integrate a mechanism for justifying aggregation of explanation metrics, or examine them in more detail, as removing sign information or grouping features can obscure a feature’s sign of influence in regression where direction can matter as much as magnitude.

  5. Develop an XAI interface that translates complex attribution maps into concise, actionable summaries for operational users, ensuring the explanations are practically useful and match the needs of downstream stakeholders.

Sources

Related papers