{"id":"65ab6180-e1c2-4f63-bfc9-25920f7be1e7","arxiv_id":"2607.24569","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"In actuated 2D wakes, POD latent dynamics forecast more reliably over long horizons than more compact CAE/VAE latents, revealing a compactness–predictability trade-off for control-oriented ROMs.","lead":"For controlled wake flows, compact nonlinear autoencoders beat POD at compression but lose on long-horizon forecasts; smoother POD latents are more predictable. That trade-off matters for model-predictive and learning-based flow control that depend on stable forecasts.","discovery_kind":"extension","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"The forecast gap is attributed to encoder-induced broadband latent dynamics, but temporal-predictor capacity is never swept — the one confound that could flip the design rule survives only as a §4 caveat.","rationale":"The reader's weakest_assumption is exactly the load-bearing point: fixing the predictor family to modest-capacity LSTMs means the POD-over-CAE forecast advantage cannot yet be attributed to latent-dynamics structure rather than to LSTM under-capacity on broadband signals. I agree with that identification and with the CONDITIONAL verdict; my pass confirms rather than revises it. Credit where due: the paper does substantial work to close alternative confounds — 20-seed encoder–predictor ensembles, latent-dimension sweep (A.2) showing larger CAE latents don't recover long-horizon accuracy, encoder-architecture sweep (A.3), sampling-rate sensitivity (A.1), phase/amplitude decomposition (Fig. 13), and an honest treatment of the divergence-rate diagnostic including the admission that truck POD λref is not quantitatively comparable. Test-set hyperparameter tuning (§2.3) is a genuine methodological flaw, but it inflates absolute E values for all branches rather than plausibly reversing a ranking that is consistent across two flows, three predictors, and four encoders, so it is not the load-bearing issue. The truck case's small encoder separation (a few E points, with seed variability dominating beyond τ=50) means the strong effect rests mostly on the pinball — a generality limit the paper itself states. The capacity confound is the one place where the causal story could actually collapse, and the paper concedes it in §4. Because the deployment context (embedded real-time MPC) makes LSTM-class predictors a defensible scoping, the result stands as subfield guidance under that scope — hence CONDITIONAL remains correct rather than REJECT (the evidence within scope is solid) or ACCEPT (the general design rule is not yet established). The proposed capacity-ladder test directly settles which side of the condition the work lands on.","tokens_in":29152,"tokens_out":5745,"duration_ms":199582,"concrete_test":"On the fluidic pinball (the discriminating case): split the 50k training snapshots into train/validation, and retrain the Direct-sequence predictor on CAE (d=10) and POD (d=22) latents across a capacity ladder — LSTM hidden/MLP dims ×1, ×4, ×16 of baseline, plus one small causal transformer (~10× parameters) — tuning on validation only and evaluating once on the untouched 20k test snapshots with the same 20-seed protocol. Check (a) whether CAE-latent training/validation loss plateaus well above the larger models' floor (direct evidence of under-capacity), and (b) whether the τ=5 median-E gap (POD 86.6% vs CAE 79.7%, Fig. 18) closes below ~2 points as capacity grows. Gap persisting >5 points across the ladder supports the encoder-dynamics claim; the gap closing at real-time-feasible sizes makes the rule LSTM-specific and requires narrowing the abstract.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim is causal: CAE/β-VAE/DKL-VAE latents are irregular and broadband, therefore LSTM forecasts on them degrade faster than on POD latents, so latent-dynamics stability can outweigh compression. For that attribution to hold, the gap must persist when the temporal model is given enough capacity to fit the harder CAE signals. That condition is never tested. §2.3 fixes the predictor family to LSTMs for real-time feasibility and a controlled comparison; Table 3 uses small networks (LSTM hidden 16–128, MLP 16–256) chosen per architecture; Appendix A sweeps sampling rate (A.1), latent dimension (A.2), and encoder architecture (A.3), but predictor capacity appears in no sweep. The only non-LSTM predictor, SINDy (App. B), is a sparse polynomial model without closure that diverges on both encoders within a few convective times — it cannot discriminate \"hard latent dynamics\" from \"wrong function class.\" Two observations sharpen the concern: (i) for the pinball, the encoded-reference divergence rates λref are comparable across all four encoders (≈0.045–0.057, Fig. 14), so the CAE latent flow is not intrinsically more sensitive — the failure is specifically a learnability/capacity interaction with the LSTM, which is precisely the reader's flagged assumption; (ii) §4 itself concedes \"more expressive sequence models could better accommodate the broadband AE latent dynamics, shifting the trade-off in favour of nonlinear encoders.\" If that holds at capacities still feasible for real-time MPC, the design rule is an LSTM-regime result, not an encoder result, and the abstract's general guidance overreaches. Test-set-tuned hyperparameters (§2.3) compound this: each model is shown at its test-optimal point, which can mask under-capacity in the AE branch while inflating absolute E values. Nothing here shows the claim is wrong — the 20-seed, multi-architecture, two-flow consistency is real — but the mechanism the headline depends on is asserted, not isolated.","agreement_with_reader":"agree"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The manuscript studies how the choice of spatial encoder affects the long-horizon predictability of latent-space reduced-order models for actuated wake flows. Two DNS datasets (a simplified 2D truck wake at Re=500 with single-input synthetic-jet actuation, and the fluidic pinball at Re=150 with three-input cylinder rotation) are compressed with POD, a convolutional autoencoder (CAE), a β-VAE, and a decomposed-KL VAE (DKL-VAE), and the latent dynamics are forecast with three LSTM-based predictors (single-step, direct sequence, derivative-based). The central empirical finding is a compactness–forecast trade-off: the nonlinear encoders achieve higher compression (e.g., d=10 vs d=22 for the pinball at matched reconstruction energy) and sharper short-term reconstructions, but their latent trajectories are irregular and broadband, so LSTM forecasts on them degrade faster and diverge more often than on POD latents. The claim is supported by end-to-end energy metrics E(τ) over 20-seed ensembles, violin distributions exposing catastrophic-failure tails for CAE-based models, phase/amplitude error decomposition, Poincaré sections, Rosenstein-style short-horizon divergence rates, and appendices sweeping sampling rate, latent dimension, encoder architecture, and a SINDy alternative predictor.","tokens_in":29602,"tokens_out":3338,"duration_ms":119402,"significance":"If the result holds, it provides practically useful, falsifiable design guidance for forecast-based flow control (MPC, model-based RL): maximal compression can be counterproductive when the downstream temporal model must extrapolate the latent dynamics. The work is a careful empirical bake-off rather than a derivation, and it ships several features that deserve explicit credit: 20-seed retraining ensembles per encoder–predictor pair (so the comparison is robust to training stochasticity), an honest clipping of the reported truck horizon at τ=50 once run-to-run variability dominates, a controlled variational-regularisation test (β-VAE and DKL-VAE) showing the trade-off is not an artefact of one penalty form, and sensitivity appendices (sampling rate, latent dimension, filter count, activation) that preserve the encoder ranking. The phase/amplitude decomposition showing phase-drift-dominated long-horizon error is a nice diagnostic. The scope is limited to two 2D low-Re DNS cases with LSTM predictors, which the authors acknowledge; within that scope the evidence is solid.","major_comments":[{"comment":"The headline causal attribution — that the forecast gap arises from encoder-induced irregular, broadband latent dynamics — is confounded by temporal-predictor capacity, which is never swept. §2.3 fixes the predictor family to LSTMs for a controlled comparison, and Appendix A sweeps sampling rate (A.1), latent dimension (A.2), and encoder architecture (A.3), but LSTM hidden size/layers appear in no sweep. Two internal observations make this load-bearing rather than pedantic: (i) Fig. 14 shows the encoded-reference divergence rates λref for the pinball are comparable across all four encoders (≈0.042–0.057), so the CAE latent flow is not intrinsically more sensitive to perturbations — the failure is a learnability/capacity interaction between the LSTM and the broadband latent signal, which is exactly what the 'latent-space stability' narrative (§3.3) attributes to the encoder alone; (ii) §4","section":"§2.3, §3.3, Fig. 14, §4"},{"comment":"The caption of Table 3 states that LSTM hyperparameters 'have been tuned via a manual hyperparameter search on the test set metrics.' The same test set is then used for the headline E(τ) curves and violin distributions (Figs. 7–8) that carry the paper's central conclusion. Tuning on the test set optimistically biases all reported accuracies and, more importantly for a comparative study, makes the fairness of the encoder comparison depend on equal tuning effort per architecture rather than on a held-out protocol. The 20-seed ensembles control training stochasticity but not this selection bias. This should be fixed by tuning on a validation split carved from the training data and reporting test metrics untouched; at minimum the authors must state how many configurations were tried per architecture and demonstrate that the encoder ranking is insensitive to the tuning protocol.","section":"§2.3, Table 3"},{"comment":"The headline comparison for the pinball contrasts POD at d=22 with CAE/β-VAE/DKL-VAE at d=10, so 'compactness versus forecast accuracy' conflates two variables: encoder class and latent dimension. Appendix A.2 sweeps d (POD d=13/22/36; CAE d=7/10/15) and reports the ranking is preserved, which addresses much of the concern, but the main-text narrative (§3.1–§3.2) should integrate this result rather than deferring it, since a reader of Figs. 7–8 alone cannot tell whether the POD advantage is a d-effect. Relatedly, a POD model truncated to d=10 (matched compactness, lower energy) would be a natural and cheap addition to Fig. 17: if d=10 POD still out-forecasts d=10 CAE, the case for latent-dynamics regularity over compression is substantially strengthened.","section":"§3.1–§3.2, Appendix A.2, Figs. 7–8, 17"}],"minor_comments":[{"comment":"Eq. (2.13): the control-sequence indices are inconsistent — the left argument lists [b_{t−L},…] while the latent lookback has L states [z_{t−L+1},…,z_t]; please harmonise the lookback indexing.","section":"§2.3, Eq. (2.13)"},{"comment":"Fig. 6 'Predicted' columns show a single representative trajectory; please state which seed/window is shown and confirm it is median-typical rather than best-case, given the heavy CAE failure tails in Figs. 7–8.","section":"Fig. 6"},{"comment":"Typos/style: 'Corresponding author:' missing space after colon (p. 2); 'an multi-layer perceptron' → 'a multi-layer perceptron' (Fig. 5 caption); several citations have collapsed spacing ('McNallyet al.2015', 'Ceruttiet al.2020', 'Solera-Ricoet al.2024', etc.) throughout.","section":"General"},{"comment":"§3.3, Fig. 14: the truck POD panel is acknowledged to lack a log-linear region, yet λref=0.075 is annotated next to values ~0.04 for the other encoders; consider omitting the annotated slope for that panel or greying it out to avoid inviting the quantitative comparison the text warns against.","section":"Fig. 14"},{"comment":"The DKL-VAE regularisation weights (λ_TC differing by 10× between truck and pinball, Table 2) are stated without a tuning protocol; a sentence on how they were selected would help reproducibility, especially since DKL-VAE is presented as a controlled test of selective regularisation.","section":"§2.2.4, Table 2"},{"comment":"Data availability states datasets and code 'will be made openly available… upon publication'; given the reproducibility emphasis (20-seed protocol), depositing at least the trained-model configs and analysis scripts at revision would strengthen the submission.","section":"Data availability"}],"recommendation":"major_revision","confidential_remarks":"The methodological core is sound and the sensitivity appendices are unusually thorough for this genre. Two items need real work before acceptance: the test-set hyperparameter tuning (Table 3) and the absence of any predictor-capacity control for the central causal claim. Both are fixable within the existing framework and datasets, so major revision rather than rejection. I also note the citation pattern leans heavily on the group's own prior pipeline (the CAE/β-VAE method, the truck case, the MPC framing all originate in the authors' earlier papers), which is natural for a follow-up study but worth the editor's awareness when assessing novelty: the incremental contribution here is the controlled comparison and the DKL-VAE test, not the framework itself."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The one thing worth knowing is the controlled finding: on a truck wake and the fluidic pinball, CAE/β-VAE/DKL-VAE compress better and look sharper short-term, but their broadband latent trajectories make LSTM forecasts degrade and diverge more often than POD. For forecast-based control (MPC/RL plant models) they argue latent predictability can beat maximal compression. That is a real, usable result inside the subfield, not a theory breakthrough.\n\nWhat they do well is the bake-off itself. Four encoders, three LSTM architectures, end-to-end energy E(τ), 20-seed ensembles, violin tails that actually show CAE catastrophic failures, plus phase portraits, Poincaré sections, phase/amplitude error, and short-horizon divergence rates. Appendices check sampling rate, latent dimension, and CAE architecture; SINDy is tried and fails cleanly. They are honest about 2D low-Re limits and the lack of a closed loop. Citations sit in the right neighborhood (Fresca, Fukami/Murata, their own prior β-VAE/truck work) without looking like a closed club.\n\nThe soft spot that matters is the one the stress note flags. They fix the temporal family to small LSTMs for “controlled comparison” and hardware feasibility, never sweep predictor capacity, and §4 admits a stronger sequence model might flip the trade-off toward AEs. Fig. 14 even shows encoded-reference divergence rates of similar magnitude across encoders on the pinball, so the gap looks like learnability/capacity interaction, not intrinsically wilder AE dynamics. Test-set-tuned hyperparameters make that a bit worse. I would not call the central claim wrong — the multi-seed, two-flow consistency is there — but the abstract’s design rule is an LSTM-regime result until someone capacity-sweeps the predictor. Minor relative to the evidence they do ship: no closed loop, data/code promised later, regime is gentle.\n\nWho it is for: people building actuation-aware ROMs for real-time control who need a concrete encoder choice under modest compute. Worth a serious referee. I would bring it to reading group if we are in AFC/ROM mode; I would cite it when choosing POD vs AE for a controlled-wake plant model, with the capacity caveat attached.","headline":"Solid empirical bake-off: for LSTM plant models on two actuated 2D wakes, POD latents beat more compact AE latents on long-horizon reliability — useful design guidance, with one untested confound on predictor capacity.","tokens_in":28909,"tokens_out":586,"would_cite":true,"duration_ms":18642,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["47.85.L-","47.32.Ff","07.05.Mh"],"model":"grok-4.5","headline":"For controlled wake flows, smoother POD latent dynamics beat more compact autoencoder latents on long-horizon forecasts.","keywords":["reduced-order models","active flow control","proper orthogonal decomposition","convolutional autoencoders","latent-space prediction","LSTM","wake flows","model predictive control"],"falsifier":"Retrain the same wakes with a higher-capacity real-time-feasible sequence model on identical CAE versus POD latents: if long-horizon median field accuracy and divergence rates equalize or reverse in favor of the autoencoder without violating latency or power budgets, the claim that latent stability outweighs compactness for these systems fails.","tokens_in":28598,"feed_emoji":"🌊","tokens_out":949,"duration_ms":24401,"temperature":0.7,"pith_summary":"Model-based flow control needs reduced-order models that both compress velocity fields and forecast how those reduced states evolve under actuation. This paper compares linear POD compression with nonlinear convolutional and variational autoencoders on two actuated 2D wakes (a simplified truck and the fluidic pinball), then trains LSTM predictors on each latent space. Autoencoders pack more energy into fewer coordinates and look sharper at short range, but their latent trajectories are irregular and broadband, so long-horizon forecasts degrade faster and diverge more often. POD produces smoother, nearly oscillatory latent signals that LSTMs learn and extrapolate more reliably. The practical claim is that for forecast-driven control, latent predictability can matter more than maximal compression.","feed_headline":"Smoother latents beat tighter ones for wake forecasts","feed_subtitle":"POD outlasts autoencoders on long-horizon controlled-wake prediction, guiding real-time flow control models","key_machinery":"A two-stage latent ROM: a spatial encoder (POD versus CAE, β-VAE, or DKL-VAE) maps velocity snapshots to low-dimensional coordinates, then an LSTM-based multi-time-delay predictor steps those coordinates forward under known control inputs; end-to-end accuracy is scored on decoded fields over prediction horizon τ.","core_discovery":"On two controlled 2D wake datasets, nonlinear autoencoders achieve higher compression and sharper short-term reconstructions than POD, yet POD-based LSTM models deliver more accurate and stable long-horizon forecasts because POD latent trajectories are smoother and narrower-band, while autoencoder latents are irregular and broadband and therefore harder to extrapolate without catastrophic divergence.","pith_inferences":["At higher Reynolds numbers with truly broadband turbulent wakes, the ranking may flip once linear POD needs too many modes to capture the energetically relevant content.","Control-conditioned or regime-local encoders could keep nonlinear compression while forcing latent trajectories onto narrower-band manifolds suited to receding-horizon optimisation.","Closed-loop MPC experiments that swap only the encoder under fixed compute budgets would quantify whether the longer stable horizon actually improves drag or base-pressure recovery.","Phase-error growth, not amplitude error, is the dominant failure mode highlighted here, so phase-aware losses or spectral penalties on the latent predictor are a natural next lever."],"forward_implications":["For model predictive control and forecast-based reinforcement learning on shedding wakes, prefer a slightly higher-dimensional but dynamically simpler latent basis over maximal compression.","Direct sequence and derivative-based LSTM predictors are better long-horizon choices than single-step autoregression for these latent ROMs.","Latent regularisation alone (β-VAE or decomposed-KL VAE) does not restore POD-level long-horizon reliability when the encoder is still driven by reconstruction.","Encoder selection for control-oriented ROMs should be scored on multi-horizon forecast stability under actuation, not only on reconstruction energy.","Hardware-feasible ROM design should co-select encoder and predictor rather than maximising compression first."],"fun_headline_variants":["POD latents beat autoencoders on long-horizon wake forecasts","Smoother POD trajectories outlast compact CAE latents in wakes","Tighter autoencoder latents falter sooner in controlled-wake prediction","Latent smoothness outweighs compression for stable wake ROMs","POD yields more reliable LSTM forecasts than nonlinear encoders"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The paper treats the forecast gap as mainly due to the encoder by holding the temporal model family fixed to LSTMs chosen for real-time hardware limits; if a stronger sequence model closed that gap for autoencoder latents at deployable cost, the design rule would weaken.","fun_headline_variants_meta":{"raw":{"variants":["POD latents beat autoencoders on long-horizon wake forecasts","Smoother POD trajectories outlast compact CAE latents in wakes","Tighter autoencoder latents falter sooner in controlled-wake prediction","Latent smoothness outweighs compression for stable wake ROMs","POD yields more reliable LSTM forecasts than nonlinear encoders"]},"model":"grok-4.5","effort":"low","cost_usd":0.004486,"raw_usage":{"total_tokens":1346,"prompt_tokens":843,"num_sources_used":0,"completion_tokens":71,"cost_in_usd_ticks":44864000,"prompt_tokens_details":{"text_tokens":843,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":432,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":843,"tokens_out":71,"duration_ms":7409,"temperature":1.0,"reasoning_tokens":432,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T11:30:08.688629+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Retrain the same wakes with a higher-capacity real-time-feasible sequence model on identical CAE versus POD latents: if long-horizon median field accuracy and divergence rates equalize or reverse in favor of the autoencoder without violating latency or power budgets, the claim that latent stability outweighs compactness for these systems fails.","supporting_citations":[],"review_version":1}