{"id":"87e26ea1-f50d-4306-b2bd-f0d5160a43b9","arxiv_id":"2506.19679","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"ELM surrogates can replace gradient-trained deep networks for low-dimensional sequential radiative transfer emulation with far less training time and data, while the image mapping task needs a 50-model ensemble that gains training speed but loses inference speed.","lead":"This paper tests Extreme Learning Machines (ELMs), a fast, non-gradient machine learning method, as surrogates for two exoplanet simulation tasks: Venus radiative transfer and stellar spot to light curve mapping. The ELM trains orders of magnitude faster and beats a recurrent network on the first task; on the image task, a 50-ELM ensemble matches a CNN with 16x less training time but 7x slower prediction.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central sample-efficiency claim is internally inconsistent: Section 5.5 asserts a 5,000-sample ELM result that Figure 3 and Table 3 do not contain, so the headline '0.28% of data' is not firmly supported.","rationale":"The reader's weakest assumption was benchmark fairness: whether the BIRNN and CNN were fairly optimized representatives of their classes. That is a legitimate concern, and the authors themselves acknowledge in Section 5.9 that batch sizes varied across models and that GPU training times may be suboptimal. However, the single most load-bearing issue is more direct: the paper's central sample-efficiency claim is internally inconsistent. Section 5.5 claims a 5,000-sample ELM achieves a six-fold MSE reduction, while Figure 3, Table 3, and Table 4 all report 10,000 samples as the training set for the best ELM. This is not a subtle benchmark-tuning question; it is a contradiction in the reported evidence that directly supports the headline '0.28% of data' statement. Resolving this inconsistency is necessary before the efficiency claims can be accepted. The reader did flag this as one of several internal inconsistencies, so there is partial agreement. I do not see a reason to move the verdict from CONDITIONAL to REJECT or UNVERDICTED, because the underlying ELM mathematics is standard, the grid-search methodology is clearly described, and the accuracy numbers in Tables 4 and 5 are internally consistent if one uses 10,000 samples. The concern is about the reproducibility and precise magnitude of the sample-efficiency gain, not about whether the central approach is fundamentally flawed. Therefore the appropriate verdict remains CONDITIONAL, pending artifact release and the 5,000- versus 10,000-sample check.","tokens_in":24431,"tokens_out":3660,"duration_ms":34271,"concrete_test":"Obtain the ELM training code and the OASIS Venus data split (or reconstruct the split from Section 2.2.1) and train the grid-optimised ELM (Gaussian activation, normal weight initialization, 1,000 hidden neurons, alpha = 1e-9) on exactly 5,000 and 10,000 training samples using the same validation set. Record the validation MSE for both. If the 5,000-sample MSE is approximately 2.0e-5 as Section 5.5 implies, then re-plot Figure 3 with the 5,000 point and correct all '0.28%' statements to '0.14%' or clarify that 10,000 is the minimum reported. If the 5,000-sample MSE is not close to the claimed value, correct Section 5.5. The same run can also resolve the 36x versus 40x prediction-speedup discrepancy by timing predictions on both 50,000 and 768,000 samples.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline contribution is that a single ELM matches or beats the BIRNN using only 0.28% of the training data (10,000 of 3,584,000 samples). This number comes from Table 3 and Table 4, where the best ELM is trained on 10,000 samples. However, Section 5.5 states: 'an ELM trained with 5,000 samples (0.14% of the data used by the BIRNN) achieved approximately a six-fold reduction in validation MSE,' and cites Figure 3. Figure 3's x-axis begins at 10,000 samples; no 5,000-sample point is plotted, and no 5,000-sample result appears in Tables 3, 4, or the text. The abstract and Table 4 use 10,000 samples. Thus the manuscript contains two mutually incompatible statements about the minimum training set size for the headline accuracy gain. If the 5,000-sample result is real, Figure 3 and Table 3 omit the most sample-efficient configuration. If it is not, the text overstates sample efficiency by a factor of two. Either way, the central sample-efficiency claim is not reproducible from the reported figures. This is a manuscript-internal inconsistency, independent of whether the BIRNN benchmark is optimally tuned.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper evaluates Extreme Learning Machines (ELMs) as low-cost surrogate models for two exoplanet-simulation tasks: (i) emulating two-stream radiative transfer in a Venus GCM (sequential, low-dimensional inputs, large training set) and (ii) mapping stellar surface maps to light curves (image inputs, high-dimensional, small training set). The authors benchmark ELMs against a BIRNN from prior work, a custom CNN, and a three-layer DNN. They report that a single ELM trained on 10,000 samples (0.28% of the BIRNN training set) achieves lower test MSE and much lower training/prediction times than the BIRNN, and that an ensemble of 50 ELMs matches or slightly beats the CNN on the image task with 16.4x less training time but 6.9x more prediction time. The paper includes grid searches over activation functions, hidden sizes, weight initializations, and training-set sizes, and it discusses limitations in Section 5.9.","tokens_in":24693,"tokens_out":4715,"duration_ms":43118,"significance":"If the reported gains hold, the paper would provide a useful practical result: a simple, gradient-free surrogate model that can be trained on a CPU in seconds while matching or beating deep-learning benchmarks on certain exoplanet simulation tasks, with substantial sample-efficiency benefits in the low-dimensional sequential setting. Strengths include the systematic hyperparameter grid, evaluation across multiple metrics (MSE, MAE, MAPE), explicit attention to sample efficiency, and an unusually candid Limitations section. However, the headline numbers are undermined by internal inconsistencies and by comparison protocols that mix CPU and GPU, different batch sizes, and single-run benchmarks. The paper's central empirical claims are defensible in principle but need correction and stronger statistical grounding before publication.","major_comments":[{"comment":"The sample-efficiency claim is internally inconsistent. Section 5.5 states that \"an ELM trained with 5,000 samples (0.14% of the data used by the BIRNN) achieved approximately a six-fold reduction in validation MSE,\" citing Figure 3. However, Figure 3's x-axis begins at 10,000 samples, no 5,000-sample point is plotted, and Table 3 reports the Experiment 1 ELM as using 10,000 samples (0.28%). The abstract and conclusion likewise rest on 0.28%. Either the 5,000-sample result is real and the most sample-efficient configuration is omitted from the figures and tables, or the text overstates sample efficiency by a factor of two. A similar issue appears in Experiment 2: Section 5.5 says an ensemble trained on 5,000 examples (78% of the full training set) marginally outperformed the CNN, but Figure 5 shows points at 3,000, 4,000, and 6,000 samples with no 5,000-point visible. These discrepancies must be resolved because the sample-efficiency claim is a central contribution of the paper.","section":"Section 5.5 vs. Figure 3 and Table 3"},{"comment":"The abstract's training-time and prediction-time ratios do not match the paper's own reported comparison. The abstract claims a \"100,000× faster training time\" and \"40× faster prediction speed\" for Experiment 1. Section 5.6 reports a \">300,000× reduction in training time\" for CPU-to-CPU comparison, and the ratio from Table 4 (146,716.66 s / 0.45 s) is ≈326,000. The abstract's 100,000× figure matches instead the GPU-trained BIRNN time from Table D1 (46,154.8 s) divided by the CPU ELM time (0.45 s), i.e., a GPU-vs-CPU comparison. For prediction, Table 4 gives 41.78 s / 1.14 s ≈ 36.6× and Section 5.7 states \"36× reduction,\" not 40×. These inconsistencies in the headline efficiency numbers need to be corrected, and the hardware basis of every timing comparison should be stated explicitly wherever speedups are quoted.","section":"Abstract vs. Section 5.6 and Table 4"},{"comment":"The benchmark comparison does not currently support the strength of the accuracy and speed claims. The BIRNN and CNN results in Tables 4 and 5 are single runs with no standard deviation or seed variation, whereas the ELM results are reported with standard deviations over random initializations. Without a distribution of BIRNN/CNN results, the claimed MSE improvements (2.00e-5 vs. 1.28e-4; 3.4e-7 vs. 4.2e-7) cannot be assessed for statistical significance. Training-time comparisons are also confounded by different batch sizes (BIRNN 512, DNN 1024, CNN 64, DNN 8) and by the use of CPU for ELMs versus GPU for some benchmarks in the abstract's 100,000× figure. Section 5.9 acknowledges these issues, but the abstract and conclusions present the speedups as unconditional. The paper should either provide matched-condition benchmark runs with uncertainty or substantially soften the comparative claims.","section":"Tables 4 and 5; Section 5.9"},{"comment":"The measurement basis for prediction time is described inconsistently. Table 4's caption states that prediction time was measured on the validation set of 768,000 samples, while Section 5.7 says the BIRNN and ELM prediction times were \"both measured on one CPU core for 50,000 samples.\" This discrepancy affects the interpretability of the reported prediction-time ratios and must be reconciled.","section":"Section 5.7 vs. Table 4"}],"minor_comments":[{"comment":"The caption of Figures F1 and F3 contains a typo: \"acorss\" should be \"across.\"","section":"Appendix F"},{"comment":"The Data Availability statement says the data were \"generated by the methods outlined in Section 2\" but does not provide access to the data or the code. For an empirical machine-learning paper, releasing the data-generation scripts and training/evaluation code would greatly aid reproducibility; at minimum, a link to a public repository should be added if the authors can share it.","section":"Data Availability"},{"comment":"The ratios σ/μ = 0.59% (Experiment 1) and 2.68% (Experiment 2) are quoted without confidence intervals; given that these are based on 100 or 1000 random seeds, a brief statement about the number of seeds and the stability of the estimate would be helpful.","section":"Section 5.3"},{"comment":"The CNN architecture table would benefit from stating the kernel size and stride for the convolutional layers explicitly, since Appendix A2 mentions kernel size 3 but the table does not list it.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely topic and the experiments are relevant to the journal's scope, but the internal numeric inconsistencies and the mixed hardware/batch-size comparison protocol need to be fixed before the headline claims can be accepted. The absence of code/data release is also a concern for a paper whose contribution is empirical; many journals now expect reproducibility artifacts for ML papers. I would not reject outright, as the core idea is sound and the limitations are acknowledged, but the authors should be asked to correct the conflicting numbers and provide matched-condition comparisons or clearly qualified claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a straightforward empirical evaluation: can Extreme Learning Machines (ELMs) serve as cheaper surrogates for exoplanet simulations? It tests them on two tasks—radiative transfer emulation (sequential data, N >> d) and stellar-surface-to-lightcurve mapping (image data, d >> N)—and reports accuracy, training time, prediction time, and sample efficiency against BIRNN and CNN benchmarks. That application-plus-benchmark is the contribution. The headline numbers are plausible, and the authors are unusually honest about limitations in Section 5.9: they flag differing batch sizes, possibly suboptimal GPU training times, and the simplicity of the image task. I see no circularity; the benchmarks come from prior work or independent training.\n\nWhat it does well: the ELM math is standard and clearly presented, the grid searches are described in enough detail to be reproduced in principle, and the discussion of when ELMs do and don't work aligns with the data. The 100,000x training-time reduction in Experiment 1 is real—the ELM trains in under a second versus hours for the BIRNN—and the ensemble result in Experiment 2 (comparable accuracy, 16x faster training, 7x slower inference) is a fair trade-off to report.\n\nThe soft spots are real but not fatal. First, the manuscript contains an internal contradiction about sample efficiency. Section 5.5 states that an ELM trained on 5,000 samples (0.14% of the BIRNN data) achieved a six-fold reduction in validation MSE, citing Figure 3. But Figure 3's x-axis starts at 10,000 samples, and no 5,000-sample result appears in the figure, Table 3, or Table 4. The abstract and tables rest on 10,000 samples (0.28%). So either the 5,000-sample result was omitted from the figures, or the text overstates sample efficiency by a factor of two. As published, the central '0.28% of the data' claim is not reproducible from the reported plots. Second, the benchmark comparisons are single runs without uncertainty; the CNN hyperparameters were selected via a grid search the authors admit may not contain the optimum, and training times use different batch sizes and hardware. Those caveats are acknowledged in 5.9, but they mean the exact speed-up factors (100,000x, 16.4x, 6.9x) should be treated as indicative, not precise. Third, there are smaller inconsistencies: the abstract says 40x faster prediction while Section 5.7 says 36x, and Figure F1's caption is mislabeled as Experiment 2 when the text describes it as Experiment 1. No code or data are released, which makes independent checking harder.\n\nWho this is for: practitioners looking for a lightweight surrogate for low-dimensional radiative transfer emulation will find the Experiment 1 results worth trying. The paper deserves a serious referee; the internal inconsistency needs to be fixed and the artifacts released, but the core empirical finding is solid enough to engage with.","headline":"A useful empirical benchmark of ELMs on two exoplanet emulation tasks, but the headline sample-efficiency claim is undercut by an internal inconsistency: the text reports a 5,000-sample result that the figures and tables don't contain.","tokens_in":903,"tokens_out":1961,"would_cite":true,"duration_ms":29458,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single extreme learning machine trained on 10,000 samples—0.28% of the benchmark's data—beats a bi-directional recurrent neural network at emulating Venus radiative transfer while training roughly 100,000 times faster.","keywords":["extreme learning machines","surrogate modeling","radiative transfer emulation","exoplanet atmospheres","sample efficiency","ensemble learning","light curves","climate model emulation"],"falsifier":"Retrain the same BIRNN architecture on the full 3,584,000-sample OASIS set with a GPU-optimal batch size (the paper flags its own 512-sample choice as likely suboptimal) and validation-based early stopping, and compare its test MSE with the 10,000-sample ELM's $(2.00 \\pm 0.01) \\times 10^{-5}$; if the re-tuned network matches or beats that error, the accuracy and sample-efficiency claims collapse even though the training-time advantage stands, and an analogous retraining of the CNN with batch size 32–64 would test whether the ensemble's 20% edge is likewise a benchmark-tuning artefact.","tokens_in":24210,"feed_emoji":"🪐","tokens_out":17757,"duration_ms":149466,"temperature":0.7,"pith_summary":"This paper argues that extreme learning machines (ELMs)—neural networks whose hidden-layer weights are drawn at random and frozen, with only the output weights solved by a single least-squares step—can replace gradient-trained deep networks as surrogate models for exoplanet simulations at a fraction of the compute cost. In the first test case, emulating two-stream radiative transfer inside a Venus climate model, a single ELM trained on 10,000 samples (0.28% of the data used by the benchmark bi-directional recurrent neural network) achieves lower test error while training roughly 100,000 times faster. In the second test case, converting stellar surface maps into light curves, a single ELM is not competitive with a convolutional network, but an ensemble of 50 ELMs slightly beats the CNN's accuracy with 16.4 times less training time and 78% of the CNN's data, at the cost of 6.9 times slower predictions. The paper's claim is that ELMs are not universally superior but context-dependent: they shine on low-dimensional sequential data with many samples, and become competitive on high-dimensional image data only when averaged into ensembles.","feed_headline":"Trains 100,000x faster, out-predicts the deep net on Venus fluxes","feed_subtitle":"A single ELM trained on 0.28% of the data beats the RNN benchmark emulating Venus radiative transfer.","key_machinery":"The paper's central object is the extreme learning machine: a single hidden-layer feedforward network in which the input-to-hidden weights $W$ and biases $b$ are sampled from a fixed distribution and never updated, the hidden activation matrix is $H = g(XW^\\top + b)$, and the only fitted parameters are the output weights $\\beta = (H^\\top H + \\alpha I)^{-1} H^\\top \\hat{Y}$, a Tikhonov-regularised least-squares solve with $\\alpha = 10^{-9}$. This closed-form solve is the whole engine of the speed gain: because the loss is a convex least-squares problem in $\\beta$, training is one matrix inversion-like step instead of thousands of gradient-descent iterations, which is why an ELM fits in 0.45 s on one CPU core while a recurrent network needs days. The second mechanism is ensembling: for the image task, diversity is created solely by drawing the frozen weights with different random seeds, and averaging 50 such predictors recovers the accuracy a single ELM lacks. A matched 15-predictor comparison shows both CNN and ELM ensembles improve with size, with ELM ensembles improving faster early on but never matching the CNN predictor-for-predictor.","core_discovery":"The central discovery is that a network that never backpropagates can match or beat purpose-built deep-learning surrogates in exoplanet simulation tasks, provided the data regime suits it. For the radiative-transfer task, the grid-optimised ELM (1,000 hidden neurons, Gaussian activation, normally distributed frozen weights) trained on 10,000 samples reaches a test MSE of $(2.00 \\pm 0.01) \\times 10^{-5}$, versus $1.28 \\times 10^{-4}$ for the BIRNN trained on 3,584,000 samples, in 0.45 seconds of single-CPU training time against 146,716 seconds, with prediction time dropping from 41.78 s to 1.14 s on the 768,000-sample validation set. For the image task, the best single ELM is an order of magnitude worse than the CNN, but averaging 50 ELMs that differ only in the random seed of their frozen weights yields a test MSE of $3.4 \\times 10^{-7}$, about 20% below the CNN's $4.2 \\times 10^{-7}$, while cutting training time from 151,414 s to 9,216 s and raising prediction time from 393 s to 2,694 s. The authors explain the contrast through dimensionality and data structure: when the sample count far exceeds the input dimension and the geometry is sequential, random projections capture the relevant structure, so the closed-form least-squares fit is enough; when the input is a high-dimensional image whose spatial layout is destroyed by flattening, many random projections must be averaged to approach convolutional performance.","pith_inferences":["A deployment the authors do not spell out is a continuously retrained radiation scheme inside the OASIS model itself: with sub-second training and inference on one CPU core, the ELM could be refit every simulation epoch, turning model drift from a hazard into a non-issue.","Because the paper fixes the ELM training set at 10,000 samples yet shows validation MSE still falling through 70,000, mapping where the error curve actually flattens would pin down the true minimum simulation budget for this task—a concrete stopping rule the paper leaves implicit.","The paper's own suggestion that the CNN would pull ahead on more realistic images is directly testable: rerun Experiment 2 with PAStar maps that include faculae and a range of temperature contrasts, and check whether the 50-ELM ensemble's 20% edge survives.","The matched 15-predictor comparison implies that a small ELM ensemble would also improve the already-winning Experiment 1 result, since the paper shows ELM ensembles gain fastest with the first few members; the authors did not run that test because the single ELM had already won."],"forward_implications":["For low-dimensional sequential emulation tasks like atmospheric radiative transfer, ELM surrogates can cut the required simulation data by more than 99% and training time by five orders of magnitude, bringing surrogate training onto a single CPU core for research groups without GPU clusters.","For image-based emulation, an ensemble of 50 ELMs offers a 16.4-fold reduction in training time with roughly 20% lower test error than the CNN, but with 6.9-fold slower predictions; the trade favours ELMs when the training budget dominates and predictions can be parallelised across CPU cores.","ELM viability tracks the ratio of sample count to input dimensionality and whether the data's geometry survives a random projection: flattening image data disrupts spatial dependencies, which is why single ELMs fail on images and succeed on sequential profiles.","Because ELM accuracy varies little with the random seed (0.59% coefficient of variation in Experiment 1, 2.68% in Experiment 2), a handful of initialisations, or a small ensemble, is a cheap route to stable predictions.","The minimal training cost enables integrating surrogate training directly into simulation runtime and makes retraining cheap when model or data drift occurs, as the paper's conclusions note."],"supporting_citations":[{"why":"Foundational formulation of the extreme learning machine; supplies the random, frozen hidden layer and analytic least-squares output fitting on which the entire method rests.","marker":"Huang et al. (2006)"},{"why":"Prior paper that supplies the OASIS Venus radiative-transfer emulation task, the BIRNN benchmark architecture and training setup, and the data preprocessing used in Experiment 1.","marker":"Tahseen et al. (2024)"},{"why":"Comparative study identifying the BIRNN as the strongest architecture for two-stream radiative-transfer emulation; motivates the Experiment 1 benchmark choice and the DNN baseline design.","marker":"Ukkonen (2022)"},{"why":"Describes the OASIS global circulation model for Venus, the simulation that generates the Experiment 1 training data.","marker":"Mendonça & Buchhave (2020)"},{"why":"Details the two-stream radiative-transfer scheme inside OASIS, the physical model the Experiment 1 surrogates emulate.","marker":"Mendonca et al. (2015)"},{"why":"The PAStar simulator that generates the stellar surface maps and light curves forming the Experiment 2 dataset.","marker":"Petralia et al. (2024)"},{"why":"Generalisation bounds for random-feature models; cited to explain why the 5,000-neuron ELM gains nothing over 2,500 and why few random features suffice.","marker":"Rudi & Rosasco (2017)"},{"why":"Prior empirical study concluding randomised networks beat backpropagation-trained networks on small datasets but lag on image tasks; frames the paper's expectations for both experiments.","marker":"Markowska-Kaczmar & Kosturek (2021)"},{"why":"AlexNet provides the convolutional design template from which the Experiment 2 CNN benchmark is adapted.","marker":"Krizhevsky et al. (2012)"}],"fun_headline_variants":["ELM surrogates beat deep nets: 100,000x faster training","Extreme learning machines outpace deep learning for exoplanet models","No backprop? ELMs beat deep nets on 0.28% of data","Lightweight ELMs: 100,000x speedup, better accuracy on flux","Ensemble ELMs rival CNNs at 16.4x faster training"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the benchmark deep networks are competently tuned representatives of their model classes: the BIRNN was taken from the authors' earlier work without re-tuning, the CNN's grid search is conceded to possibly miss the optimum, and batch sizes and CPU/GPU conditions differed across models, so if the benchmarks are under-tuned, the headline speed and accuracy gaps are overstated.","fun_headline_variants_meta":{"raw":{"variants":["ELM surrogates beat deep nets: 100,000x faster training","Extreme learning machines outpace deep learning for exoplanet models","No backprop? ELMs beat deep nets on 0.28% of data","Lightweight ELMs: 100,000x speedup, better accuracy on flux","Ensemble ELMs rival CNNs at 16.4x faster training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0007,"raw_usage":{"total_tokens":3301,"prompt_tokens":1225,"completion_tokens":2076,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":841,"completion_tokens_details":{"reasoning_tokens":1972}},"tokens_in":841,"tokens_out":2076,"duration_ms":15069,"temperature":1.0,"reasoning_tokens":1972,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:28:38.056396+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the same BIRNN architecture on the full 3,584,000-sample OASIS set with a GPU-optimal batch size (the paper flags its own 512-sample choice as likely suboptimal) and validation-based early stopping, and compare its test MSE with the 10,000-sample ELM's $(2.00 \\pm 0.01) \\times 10^{-5}$; if the re-tuned network matches or beats that error, the accuracy and sample-efficiency claims collapse even though the training-time advantage stands, and an analogous retraining of the CNN with batch size 32–64 would test whether the ensemble's 20% edge is likewise a benchmark-tuning artefact.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior paper that supplies the OASIS Venus radiative-transfer emulation task, the BIRNN benchmark architecture and training setup, and the data preprocessing used in Experiment 1."},{"cited_title":"Exploring Pathways to More Accurate Machine Learning Emulation of Atmospheric Radiative Transfer , Journal of Advances in Modeling Earth Systems\\/ , 14 (4), e2021MS002875","cited_arxiv_id":null,"evidence_quote":"Comparative study identifying the BIRNN as the strongest architecture for two-stream radiative-transfer emulation; motivates the Experiment 1 benchmark choice and the DNN baseline design."},{"cited_title":"M., Read, P","cited_arxiv_id":null,"evidence_quote":"Details the two-stream radiative-transfer scheme inside OASIS, the physical model the Experiment 1 surrogates emulate."},{"cited_title":"PAStar: a model for stellar surface from the Sun to active stars","cited_arxiv_id":"2412.10035","evidence_quote":"The PAStar simulator that generates the stellar surface maps and light curves forming the Experiment 2 dataset."},{"cited_title":"& Rosasco, L., 2017","cited_arxiv_id":null,"evidence_quote":"Generalisation bounds for random-feature models; cited to explain why the 5,000-neuron ELM gains nothing over 2,500 and why few random features suffice."},{"cited_title":"& Kosturek, M., 2021","cited_arxiv_id":null,"evidence_quote":"Prior empirical study concluding randomised networks beat backpropagation-trained networks on small datasets but lag on image tasks; frames the paper's expectations for both experiments."},{"cited_title":"E., 2012","cited_arxiv_id":null,"evidence_quote":"AlexNet provides the convolutional design template from which the Experiment 2 CNN benchmark is adapted."}],"review_version":2}