Interpreting CFD Surrogates through Sparse Autoencoders

Shusen Liu; Yeping Hu

REVIEW 5 major objections 6 minor 21 references

Interpreting CFD Surrogates through Sparse Autoencoders

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Training a sparse autoencoder on a frozen CFD surrogate's node embeddings surfaces vorticity-aligned features that beat PCA and embedding-norm baselines at highlighting flow structures.

desk verdict Credible first transfer of SAEs to CFD surrogate interpretability, but the quantitative eval is unverifiable as written and the disentanglement/trustworthiness claims outrun the evidence. read the letter →

arxiv 2507.16069 v1 pith:5D7SKVSC submitted 2025-07-21 cs.CE cs.LG

Yeping Hu , Shusen Liu This is my paper · ORCID

classification cs.CEcs.LG

keywords sparseautoencodersdictionarylearninginterpretabilityCFDsurrogatemodelsgraphneuralnetworksMeshNetvorticityalignmentpost-hocexplanation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that sparse autoencoders—a dictionary-learning technique from language-model interpretability—can make graph-based CFD surrogates interpretable after the fact. An autoencoder is trained on the frozen node embeddings of a pretrained MeshGraphNet cylinder-flow surrogate, and the resulting sparse dictionary separates into latent dimensions whose individual spatial footprints match recognizable flow physics such as boundary-layer separation, wall effects, and wake vortices. The authors claim this is a model-agnostic route to trustworthiness because the surrogate itself is never altered or retrained. If correct, CFD practitioners would gain a cheap post-hoc procedure for inspecting what a black-box surrogate's internal representation actually tracks, with quantitative saliency that beats embedding-norm, PCA, and random baselines on vortex-region alignment.

What carries the argument

The load-bearing object is a sparse autoencoder trained on the frozen node embeddings $\{h^L_{i,t}\}$ of the pretrained surrogate. The encoder is a linear map followed by a ReLU, $z = \sigma((h - b_{\mathrm{dec}}) W_{\mathrm{enc}} + b_{\mathrm{enc}})$ with $W_{\mathrm{enc}} \in \mathbb{R}^{d_{\mathrm{in}} \times d_{\mathrm{hid}}}$ and expansion factor $\kappa = 8$; the decoder reconstructs via $\hat h = z W_{\mathrm{dec}} + b_{\mathrm{dec}}$ with $W_{\mathrm{dec}} \in \mathbb{R}^{d_{\mathrm{hid}} \times d_{\mathrm{in}}}$, whose rows are renormalized to unit $\ell^2$ norm after each step. The objective $L_{\mathrm{SAE}} = \|\hat h - h\|_2^2 + \lambda \|z\|_1$ balances reconstruction fidelity against sparsity, and the normalized decoder rows constitute the feature dictionary. Relevance scores—variance, mean absolute, and entropy—rank the dictionary dimensions; taking the top $K$ globally or per time step and summing each node's activations gives mesh-space saliency maps $a_{i,t} = \sum_{d \in K_t} z^{(d)}_{i,t}$, linking abstract latent coordinates back to spatial flow structures.

What would settle it

Recompute Table 2 with an explicitly specified vorticity estimator, such as the discrete curl of the surrogate's predicted velocity field, and with alternative physical masks (pressure gradient, enstrophy, boundary-layer distance); if SAE saliency no longer outperforms PCA or embedding-norm under any of these masks, the paper's claim that SAE features align with physics would lose its quantitative support.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the 128-dimensional node embeddings of a trained MeshGraphNet are not irreducibly entangled: an overcomplete sparse autoencoder with an eightfold expansion factor ($d_{\mathrm{hid}} = 1024$) and an L1 sparsity penalty learns decoder rows that act as individual feature detectors, each firing at a distinct mesh region. Aggregating the top salient dimensions—ranked by variance, mean absolute activation, or entropy—produces a node-level saliency map, and the first two rankings match the top-10% high-vorticity nodes better than embedding-norm, PCA, or random baselines (F1 0.60 versus 0.55, 0.49, and 0.09). The paper takes this as evidence that the surrogate's latent space contains disentangled, physically meaningful concepts that can be extracted post-hoc without modifying the predictive model.

Load-bearing premise

The quantitative ranking of all methods assumes that the top 10% vorticity nodes, computed by a definition the paper leaves unresolved in a dangling citation, are the correct ground truth for what an interpretable feature should highlight.

Editorial extensions

If this is right

A frozen graph surrogate's node embeddings can be decomposed into a sparse dictionary whose individual atoms localize to distinct flow features, so interpretability does not have to be engineered into the architecture.
SAE-based saliency with variance or mean-absolute scoring beats embedding-norm, PCA, and random baselines at highlighting high-vorticity nodes, providing a quantitative post-hoc trust signal.
Because global and time-local top-$K$ selections oscillate with the vortex-shedding period (measured by Jaccard similarity), the same dictionary can separate persistent flow structures from transient shedding events without retraining.
The pipeline is agnostic to the scoring criterion's granularity: variance and mean-abs track coherent structures while entropy tracks short-lived fine-scale features, so users can tune the explanation's temporal resolution.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural next test the paper does not run is causal: amplify or suppress a single dictionary atom in the frozen surrogate and check whether the predicted flow field changes most in that atom's region; positive results would upgrade the dictionary from correlational to causal interpretability.
The time series of individual atom activations could itself be used to detect vortex-shedding events or regime changes automatically, turning the SAE into a flow-diagnostics tool rather than only a saliency map.
The quantitative evaluation targets only vorticity; extending the same comparison to pressure-gradient, shear, or boundary-layer masks would show whether SAE features are broadly physical or specifically vorticity-tuned.
Training the same SAE pipeline on a different surrogate or mesh geometry would test the claimed model-agnosticism more strongly than the single CylinderFlow dataset used here.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: this is the first application of SAEs to post-hoc interpretation of graph-based CFD surrogates, and the qualitative spatial visuals are suggestive. The quantitative evaluation, though, is unverifiable as written—the vorticity ground truth in Section 5.4 is cited as '[?]'—and the central claims about monosemanticity and trustworthiness outrun the evidence.

What is actually new: the SAE architecture and loss are directly from Gao et al. 2024, but applying it to frozen MeshGraphNet node embeddings is new, as is the global-versus-time-local saliency distinction. The mesh-space projections, especially the individual latent dimensions in Figure 3, show spatially disjoint footprints that look like genuinely distinct flow features. That is real evidence of disentanglement, at least at a qualitative level.

The soft spots are concentrated in Section 5.4. The missing citation is not a typo; Table 2 is meaningless until the exact vorticity mask is defined (discretization, filtering, threshold). Even with a proper reference, spatial overlap with high-vorticity nodes is a weak proxy. A mixed, non-monosemantic feature can still concentrate on vorticity, and a surrogate that focuses on vorticity can still be untrustworthy. The paper itself notes in Section 3.2 that the construction enables causal ablations, but none are run. The identical rounded scores for SAE(variance) and SAE(mean_abs), and for SAE(entropy) and Embedding-norm, suggest the comparison lacks sensitivity; there are no error bars or significance tests, and only one dataset and one surrogate configuration are tested. The circularity concern from the reader is minor—selection is by variance/entropy, not the target—so that is not the problem.

Who this is for: people working on XAI for learned physics simulators. The qualitative pipeline and the temporal/saliency idea are worth a read. But the abstract's claims need to be scaled back or supported by per-feature purity metrics and at least one causal intervention.

For peer review: send it out. The idea is timely and the flaws are fixable. Reviewers should require the missing reference, error bars, a second dataset or surrogate, and either direct evidence of monosemanticity or a more modest framing.

Referee Report

5 major / 6 minor

Summary. The paper proposes a post-hoc interpretability framework for graph-based CFD surrogates. The authors train a sparse autoencoder on frozen node embeddings from a pretrained MeshGraphNet surrogate, rank the resulting dictionary dimensions by variance, mean absolute activation, or entropy, and project the most active dimensions back to the mesh. The evaluation consists of qualitative visualizations (Figures 1-3) and a quantitative comparison (Section 5.4, Table 2) of node-level saliency against a vorticity mask, with the claim that SAE-based saliency outperforms embedding-norm, PCA, and random baselines and thereby identifies monosemantic concepts aligned with physical phenomena.

Significance. If the evaluation were fully specified and the claims appropriately scoped, the paper would make a useful contribution by demonstrating that sparse dictionary learning can be applied post-hoc to physics-based surrogate models and can yield spatially meaningful node saliency. The work has clear strengths: it uses a public benchmark (CylinderFlow), evaluates multiple ranking criteria, compares against several baselines, and its quantitative target (vorticity) is external to the surrogate, so the core comparison is not circular. However, as it stands, the evidence does not support the stronger claims about monosemanticity and trustworthiness, and the central quantitative result rests on an unresolved citation and lacks statistical grounding. The approach is promising and the missing pieces are addressable, so major revision is appropriate rather than rejection.

major comments (5)

[Section 5.4, Table 2] The ground truth for the quantitative evaluation is defined only as 'the top 10% vorticity nodes (calculated using the method described in [?])'. Because this citation is unresolved, the discrete vorticity estimator (finite-difference scheme, mesh curl, filtering, thresholding) is unspecified, so Table 2 cannot be reproduced and a different vorticity estimator could change the reported ranking. Specify the vorticity computation exactly, or replace it with a fully specified reference.
[Section 3.2 and Section 5.4] The SAE is trained on 'the collection of node embeddings {h^L_i,t} from testing data' (Section 3.2), and the evaluation in Section 5.4 appears to use the same test rollouts. If the dictionary and the top-K selection are derived from the same snapshots that are scored, the results may reflect in-sample selection rather than generalizable saliency. State the exact data split used for SAE training, top-K selection, and evaluation, and if the split differs, describe it explicitly.
[Abstract and Section 6 vs. Section 5.4] The abstract and conclusion claim that the method identifies 'monosemantic concepts' and 'enhance[s] explainability and trustworthiness', but the only quantitative result is spatial overlap of a node-level saliency map with a vorticity mask. Such overlap can be produced by mixed, non-monosemantic features and does not by itself establish trustworthiness. Since Section 3.2 states that the construction enables causal ablations, report at least one per-feature purity metric (e.g., activation sparsity per atom, footprint overlap) or a causal intervention test, or restrict the claims to spatial localization only.
[Table 2] The reported scores for SAE(variance) and SAE(mean_abs) are identical to three decimal places, and SAE(entropy) is identical to embedding-norm, which suggests the comparison has low sensitivity. No error bars, per-timestep distributions, or significance tests are given, so the reader cannot judge whether the differences between methods are meaningful. Report variability across test cases and timesteps and a statistical test (e.g., paired bootstrap or Wilcoxon) for the Table 2 comparisons.
[Section 4.3 and Section 4.5] Salient dimensions are selected by variance, mean-abs, or entropy of SAE activations, which are generic informativeness criteria unrelated to vorticity or any other physical target. Consequently, the observed vortex-region alignment (Table 2) may be a property of the node-level aggregation a_{i,t} in Section 4.5 rather than of the dictionary atoms. The paper should separately evaluate the contribution of the dictionary (vs. the aggregation rule) and, if the claim is about physical concepts, select or rank dimensions by a physics-aware criterion or justify why generic informativeness should yield physical alignment.

minor comments (6)

[Section 4.5 and Table 2] Report the actual values of η and K used for the quantitative results; the text gives only visualization budgets (η = 20, 85, 300).
[Figure 1 caption] The caption contains a typo, 'T op-50'; also define 'Top-50' as K=50 in the caption or text.
[Table 1] Table 1 uses z* and p_{b,d} without defining z*; align the notation with Eq. (4).
[Section 4.2] The SAE training description omits learning rate, number of epochs, and batch size; include these details for reproducibility.
[Introduction] The speedup claim '11x-290x' cites Beale and Majda (1985), which is a vortex-methods paper and not a surrogate-model benchmark; this citation appears inappropriate.
[Section 5.3] The 'spatial disjointness' claim is based on visual inspection of three selected dimensions; a quantitative overlap metric would strengthen the disentanglement statement.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the SAE saliency is selected by activation statistics and scored against an external vorticity mask; the unresolved ground-truth citation is a reproducibility concern, not circularity.

full rationale

The paper's central derivation is not circular. The sparse autoencoder is trained post-hoc on frozen MGN node embeddings using an ℓ1-sparsity objective (Eq. 6), and its hyperparameters (κ = 8, λ = 3×10^-4) are fixed independently of the evaluation target. The salient latent dimensions are ranked by activation statistics alone—variance, mean absolute activation, and entropy (Table 1)—not by alignment with vorticity. The quantitative benchmark in Section 5.4 then compares each method's top-η mesh nodes against an external ground-truth mask of high-vorticity nodes derived from the CFD simulation field, not from the SAE's own activations or from the surrogate's outputs. Baselines such as embedding-norm, PCA, and random node sampling are independent comparators, so Table 2 is a fair external comparison rather than a fitted-input prediction. The unresolved citation '[?]' for the vorticity computation is a genuine reproducibility gap, and the abstract's 'monosemantic' and 'trustworthiness' phrasing goes beyond what a spatial-overlap metric can establish, but these are correctness and scope concerns, not circularity. Self-citations in the introduction ([Hu et al., 2023], [Hu et al., 2024]) are illustrative context about surrogate models and CFD cost and do not carry the derivation. Accordingly, the paper earns a low circularity score despite its evaluation weaknesses.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method's outputs depend on several hand-chosen hyperparameters (kappa, lambda, K, eta, B) and on the domain assumption that high-vorticity regions are the correct ground truth. None of these are derived from theory or validated by sensitivity analysis. The paper also assumes the pretrained surrogate's embeddings are physically meaningful without reporting its prediction accuracy. This ledger is moderate: no invented entities and the benchmark is external, but the hand-chosen parameters are numerous and the proxy for interpretability is a single physical quantity.

free parameters (5)

SAE expansion factor kappa = 8
Chosen by hand; sets dictionary width to 1024 for 128-dim embeddings. No sensitivity analysis or justification is provided.
Sparsity penalty lambda = 3e-4
Chosen by hand; balances reconstruction and L1 sparsity. No search or ablation is reported.
Top-K number of salient dimensions K = 50
Used for Jaccard and vortex-alignment experiments; selected without a criterion or sensitivity study.
Sampling budget eta for top nodes = 20, 85, 300 (Fig. 2); 100 (Fig. 3)
User-defined budget for mesh-space visualization; different values change the precision/recall tradeoff but no sweep is reported in the quantitative table.
Entropy histogram bins B = not reported
Table 1 defines entropy over histogram bins but the number of bins B is not specified, affecting the entropy ranking result.

assumptions (4)

domain assumption The final-layer node embeddings h^L_i,t of the pretrained MGN contain sufficient information to recover meaningful physical concepts.
The whole pipeline analyzes only the frozen node embeddings from the last message-passing layer; if the relevant physics is encoded in edge features or intermediate layers, the dictionary is incomplete. Invoked in Section 3.2.
domain assumption Vorticity alignment is a valid proxy for interpretability and trustworthiness of a CFD surrogate.
Section 5.4 uses top-10% vorticity nodes as the only quantitative ground truth; no alternative physical quantities or human-evaluation study is used.
domain assumption The pretrained MeshGraphNets surrogate is accurate enough that its embeddings reflect the true flow physics.
The paper never reports the surrogate's prediction error on the test set; if the surrogate is poorly trained, the SAE features would reflect the model's errors rather than physics. Assumed in Section 4.2.
ad hoc to paper Sparse dictionary learning on ReLU activations yields features that are monosemantic in the embedding space.
The paper borrows this assumption from LLM interpretability literature (Cunningham et al., Gao et al.) without validating it for CFD embeddings; it claims monosemanticity without a direct measure.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Interpreting CFD Surrogates through Sparse Autoencoders." pith.science (2026). https://pith.science/paper/5D7SKVSC

@misc{pith2026250716069,
  author       = {Pith},
  title        = {Pith review of: Interpreting CFD Surrogates through Sparse Autoencoders},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5D7SKVSC}},
  note         = {Machine review of arXiv:2507.16069}
}

read the original abstract

Learning-based surrogate models have become a practical alternative to high-fidelity CFD solvers, but their latent representations remain opaque and hinder adoption in safety-critical or regulation-bound settings. This work introduces a posthoc interpretability framework for graph-based surrogate models used in computational fluid dynamics (CFD) by leveraging sparse autoencoders (SAEs). By obtaining an overcomplete basis in the node embedding space of a pretrained surrogate, the method extracts a dictionary of interpretable latent features. The approach enables the identification of monosemantic concepts aligned with physical phenomena such as vorticity or flow structures, offering a model-agnostic pathway to enhance explainability and trustworthiness in CFD applications.

Figures

Figures reproduced from arXiv: 2507.16069 by the authors.

**Figure 3.** Spatial footprint of three individual salient latent dimen [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

**Figure 2.** Mesh–space footprint of aggregated salient latent dimen [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 9 canonical work pages

[1]

Data-Driven Insights into Jet Turbulence: Explainable AI Approaches

[Amico et al., 2025] Enrico Amico, Lorenzo Matteucci, and Gioacchino Cafiero. Data-driven insights into jet tur- bulence: Explainable ai approaches. arXiv preprint arXiv:2503.02126,

work page Pith review arXiv 2025
[2]

arXiv preprint arXiv:2408.05147,

arXiv
[5]

Scaling and evaluating sparse autoencoders

[Gao et al., 2024] Leo Gao, Tom Dupr’e la Tour, Henk Tillman, Gabriel Goh, Rajan Troll, Alec Radford, Ilya Sutskever, Jan Leike, and Jeffrey Wu. Scaling and evaluating sparse autoencoders. arXiv preprint arXiv:2406.04093,

arXiv 2024
[7]

Comparative study of machine learning tech- niques for post-combustion carbon capture systems

[Hu et al., 2024] Yeping Hu, Bo Lei, Yash Girish Shah, Jose Cadena, Amar Saini, Grigorios Panagakos, and Phan Nguyen. Comparative study of machine learning tech- niques for post-combustion carbon capture systems. Fron- tiers in Artificial Intelligence, 7:1441934,

work page 2024
[8]

Gemma scope: Open sparse autoencoders everywhere all at once on gemma

[Lieberum et al., 2024] Tom Lieberum, Senthooran Raja- manoharan, Arthur Conmy, Lewis Smith, Nicolas Son- nerat, Vikrant Varma, J ´anos Kram ´ar, Anca Dragan, Ro- hin Shah, and Neel Nanda. Gemma scope: Open sparse autoencoders everywhere all at once on gemma

work page 2024
[10]

K-sparse autoencoders

[Makhzani and Frey, 2013] Alireza Makhzani and Bren- dan Frey. K-sparse autoencoders. arXiv preprint arXiv:1312.5663,

arXiv 2013
[11]

Enhancing neural network inter- pretability with feature-aligned sparse autoencoders.arXiv preprint arXiv:2411.01220,

[Marks et al., 2024] Luke Marks, Alasdair Paren, David Krueger, and Fazl Barez. Enhancing neural network inter- pretability with feature-aligned sparse autoencoders.arXiv preprint arXiv:2411.01220,

arXiv 2024
[12]

Michaud, Max Tegmark, and Christian Schroeder de Witt

[Mudide et al., 2024] Anish Mudide, Joshua Engels, Eric J. Michaud, Max Tegmark, and Christian Schroeder de Witt. Efficient dictionary learning with switch sparse autoen- coders. arXiv preprint arXiv:2410.08201,

arXiv 2024

Show all 21 references

[13]

Diab, and Virginia Smith

[Muhamed et al., 2024] Aashiq Muhamed, Mona T. Diab, and Virginia Smith. Decoding dark matter: Specialized sparse autoencoders for interpreting rare concepts in foun- dation models. arXiv preprint arXiv:2411.00743,

2024 arXiv
[14]

Uncertainty quantification and polynomial chaos techniques in computational fluid dy- namics

[Najm, 2009] Habib N Najm. Uncertainty quantification and polynomial chaos techniques in computational fluid dy- namics. Annual review of fluid mechanics , 41(1):35–52,

2009
[17]

Jumping ahead: Improving reconstruction fidelity with jumprelu sparse au- toencoders

[Rajamanoharan et al., 2024b] Senthooran Rajamanoharan, Tom Lieberum, Nicolas Sonnerat, Arthur Conmy, Vikrant Varma, J´anos Kram ´ar, and Neel Nanda. Jumping ahead: Improving reconstruction fidelity with jumprelu sparse au- toencoders. arXiv preprint arXiv:2407.14435,

arXiv
[18]

Sparse autoencoders for scien- tifically rigorous interpretation of vision models

[Stevens et al., 2025] Samuel Stevens, Wei-Lun Chao, Tanya Berger-Wolf, and Yu Su. Sparse autoencoders for scien- tifically rigorous interpretation of vision models. arXiv preprint arXiv:2502.06755,

2025
[19]

Universal sparse autoencoders: Interpretable cross-model concept alignment

[Thasarathan et al., 2025] Harrish Thasarathan, Julian Forsyth, Thomas Fel, Matthew Kowal, and Kon- stantinos Derpanis. Universal sparse autoencoders: Interpretable cross-model concept alignment. arXiv preprint arXiv:2502.03714,

2025
[20]

Artificial intelligence explainability requirements of the ai act and metrics for measuring compliance

[Walke et al., 2023] Fabian Walke, Lars Bennek, and Till J Winkler. Artificial intelligence explainability requirements of the ai act and metrics for measuring compliance. In International Conference on Wirtschaftsinformatik, pages 113–129. Springer,

2023
[21]

An explainable deep learning model based on hydrological principles for flood simulation and forecasting

[Xiang et al., 2025] Xin Xiang, Shenglian Guo, Chenglong Li, and Yun Wang. An explainable deep learning model based on hydrological principles for flood simulation and forecasting. EGUsphere [preprint], 2025

2025
[1985]

Sparse autoencoders find highly interpretable features in language models

[Cunningham et al., 2023] Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert Huben, and Lee Sharkey. Sparse autoencoders find highly interpretable features in language models. arXiv preprint arXiv:2309.08600,

2023 arXiv
[2009]

Learning mesh- based simulation with graph networks

[Pfaff et al., 2020] Tobias Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, and Peter Battaglia. Learning mesh- based simulation with graph networks. In International conference on learning representations,

2020
[2020]

Improving dictionary learning with gated sparse autoen- coders

[Rajamanoharan et al., 2024a] Senthooran Rajamanoharan, Arthur Conmy, Lewis Smith, Tom Lieberum, Vikrant Varma, J ´anos Kram ´ar, Rohin Shah, and Neel Nanda. Improving dictionary learning with gated sparse autoen- coders. arXiv preprint arXiv:2404.16014,

arXiv
[2023]

High order accurate vortex methods with explicit velocity kernels

[Beale and Majda, 1985] J Thomas Beale and Andrew Ma- jda. High order accurate vortex methods with explicit velocity kernels. Journal of Computational Physics , 58(2):188–208,

1985
[2024]

Graph learning in physical-informed mesh-reduced space for real-world dynamic systems

[Hu et al., 2023] Yeping Hu, Bo Lei, and Victor M Castillo. Graph learning in physical-informed mesh-reduced space for real-world dynamic systems. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4166–4174,

2023
[2025]

Interpretable fine-tuning for graph neural net- work surrogate models

[Barwey and Maulik, 2023] Shivam Barwey and Romit Maulik. Interpretable fine-tuning for graph neural net- work surrogate models. arXiv preprint arXiv:2311.07548,

2023 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Data-Driven Insights into Jet Turbulence: Explainable AI Approaches

[Amico et al., 2025] Enrico Amico, Lorenzo Matteucci, and Gioacchino Cafiero. Data-driven insights into jet tur- bulence: Explainable ai approaches. arXiv preprint arXiv:2503.02126,

work page Pith review arXiv 2025

[2] [2]

arXiv preprint arXiv:2408.05147,

arXiv

[3] [5]

Scaling and evaluating sparse autoencoders

[Gao et al., 2024] Leo Gao, Tom Dupr’e la Tour, Henk Tillman, Gabriel Goh, Rajan Troll, Alec Radford, Ilya Sutskever, Jan Leike, and Jeffrey Wu. Scaling and evaluating sparse autoencoders. arXiv preprint arXiv:2406.04093,

arXiv 2024

[4] [7]

Comparative study of machine learning tech- niques for post-combustion carbon capture systems

[Hu et al., 2024] Yeping Hu, Bo Lei, Yash Girish Shah, Jose Cadena, Amar Saini, Grigorios Panagakos, and Phan Nguyen. Comparative study of machine learning tech- niques for post-combustion carbon capture systems. Fron- tiers in Artificial Intelligence, 7:1441934,

work page 2024

[5] [8]

Gemma scope: Open sparse autoencoders everywhere all at once on gemma

[Lieberum et al., 2024] Tom Lieberum, Senthooran Raja- manoharan, Arthur Conmy, Lewis Smith, Nicolas Son- nerat, Vikrant Varma, J ´anos Kram ´ar, Anca Dragan, Ro- hin Shah, and Neel Nanda. Gemma scope: Open sparse autoencoders everywhere all at once on gemma

work page 2024

[6] [10]

K-sparse autoencoders

[Makhzani and Frey, 2013] Alireza Makhzani and Bren- dan Frey. K-sparse autoencoders. arXiv preprint arXiv:1312.5663,

arXiv 2013

[7] [11]

Enhancing neural network inter- pretability with feature-aligned sparse autoencoders.arXiv preprint arXiv:2411.01220,

[Marks et al., 2024] Luke Marks, Alasdair Paren, David Krueger, and Fazl Barez. Enhancing neural network inter- pretability with feature-aligned sparse autoencoders.arXiv preprint arXiv:2411.01220,

arXiv 2024

[8] [12]

Michaud, Max Tegmark, and Christian Schroeder de Witt

[Mudide et al., 2024] Anish Mudide, Joshua Engels, Eric J. Michaud, Max Tegmark, and Christian Schroeder de Witt. Efficient dictionary learning with switch sparse autoen- coders. arXiv preprint arXiv:2410.08201,

arXiv 2024