{"id":"28370493-2960-4f84-8730-16211191beee","arxiv_id":"2601.09305","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"PMoE is a mixture-of-experts RANS turbulence model that grows by adding flow-regime-specific experts and autoencoder routing modules without retraining prior modules, improving accuracy on seen and unseen cases.","lead":"A new machine-learning framework, PMoE, lets RANS turbulence models be expanded with a new 'expert' model for each new flow regime without retraining older ones. Tests across airfoil wake, channel, separated hill, and square-duct flows show better accuracy than the baseline with only a small inference-cost increase.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Router novelty detection is under-specified and internally inconsistent: reported confidences contradict the stated 99.9th-percentile threshold protocol, so the core 'detect novel flow' claim lacks a validated mechanism.","rationale":"The paper's central claim is that PMoE enables continual learning for RANS by using an autoencoder router to detect when a flow is not represented and then adding a new expert without degrading old ones. The modular design is coherent, and the forgetting-avoidance argument is structurally sound because old components are frozen. The empirical evidence is directionally positive but qualitative, and the reader already marked this CONDITIONAL. My stress-test focuses on the router's novelty-detection mechanism, which is the load-bearing part of the central claim. The internal inconsistency between the stated 99.9th-percentile threshold protocol and the reported confidence values (e.g., 91.5% for the training case PH1p0) is a concrete correctness risk: it suggests either the protocol is misdescribed or the thresholds were tuned rather than derived from the stated quantile. If the threshold is effectively tuned to clear the 90% acceptance bar, the claim that the router unsupervisedly identifies novel regimes is weaker than presented. This does not warrant rejection, because the four demonstrated regimes are well separated and the framework may still work in practice, but it does require resolution before the continual-learning claim can be considered fully supported. My proposed check—recomputing p_k under the stated protocol—would settle whether the discrepancy is a reporting error or a substantive flaw. I therefore keep the reader's CONDITIONAL verdict unchanged.","tokens_in":22230,"tokens_out":8627,"duration_ms":90040,"concrete_test":"Reimplement or run the released router code and recompute p_k for each training case using the protocol in §2.2.2 exactly: set T_k to the 99.9th percentile of reconstruction errors over the N training samples for component k, then compute p_k on those same N samples. If p_k for PH1p0 is ≈99.9% rather than the reported 91.5%, the discrepancy is a reporting error but the mechanism is sound; if p_k is ~91.5%, the stated threshold protocol is not what was used, and the novelty-detection claim must be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In §2.2.2 the router defines T_k as the 99.9th percentile of the reconstruction error of the training data, so the confidence p_k on that same training data should be ≈99.9%. Yet Table 3 reports the new component's confidence for separated flow as 93.5% and for corner flow as 97.0%, and Table 5 lists PH1p0 (the training case for C2) at 91.5% and SD5693 (a training case for C3) at 97.0%—far below 99.9%. Either the threshold is not actually set as stated, the confidence is computed on different samples, or the thresholds are tuned so that training data clears the 90% acceptance bar. The autoencoder router's known/novel decision is the linchpin of the continual-learning claim: if the threshold is arbitrary or miscalibrated, a genuinely novel regime that shares feature distributions with an existing component can be assigned to the wrong expert, and the 'progressive' expansion criterion is unreliable. The paper itself notes T_accept=90% would need tuning for future applications, but no calibration on out-of-family flows is provided, so 'detects when a flow is not represented' is not yet supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PMoE, a progressive mixture-of-experts framework for RANS turbulence modelling. A modular autoencoder router assigns each local flow state to a specialized expert; when no component achieves 90% confidence, a new autoencoder component and expert are trained on the new regime, leaving old components fixed. The framework is built in four stages using airfoil wake, channel, periodic hill, and square-duct flows, with experts designed respectively as baseline SA, a symbolic-regression damping correction, a FIML production-term correction, and a spatially varying QCR coefficient. The final PMoE-S3 model is tested on training and unseen cases (C8000, PH0p8/1.2/1.5, SD3500) and is reported to improve on the SA baseline at negligible computational overhead.","tokens_in":22608,"tokens_out":8505,"duration_ms":81224,"significance":"The idea of continual learning for RANS closures is timely and the modular architecture is well matched to the goal. The a posteriori validation on unseen Reynolds numbers and hill slopes is the strongest part of the paper: those cases were not used to train the experts or router components, so the transfer evidence is independent. Sparse activation and the reported <4% wall-clock overhead are also concrete practical benefits. However, the router's novelty-detection mechanism is not yet convincingly validated (threshold inconsistency and lack of out-of-family tests), and the accuracy claim rests solely on visual comparisons. If the authors resolve the threshold issue and add quantitative error metrics, the framework would be a useful contribution.","major_comments":[{"comment":"The stated threshold protocol is internally inconsistent. The text defines T_k as the 99.9th percentile of the training reconstruction error, so the confidence p_k on the same training data should be approximately 99.9%. However, Table 3 reports 93.5% for the newly added C2 and 97.0% for C3, and Table 5 gives 91.5% for PH1p0 (training case for C2) and 97.0% for SD5693 (training case for C3). Either the threshold is not computed as stated, the confidence is evaluated on different samples, or the threshold is tuned to pass the 90% acceptance bar. Since the known/novel decision is the linchpin of the continual-learning claim, the mechanism needs a precise, consistent specification and calibration.","section":"§2.2.2, Eq. (2.6), Tables 3 and 5"},{"comment":"The 'confidence distribution' is inconsistent with Eq. (2.6). Under that definition, p_k is a per-case scalar and a point can be recognized by several components, so percentages need not sum to 100. Table 5 presents a hard per-point assignment over components plus an 'Unknown' column, which is a different routing rule not defined in the text. Please clarify how the entries are computed and how the '>90% classification accuracy' claim follows from Eq. (2.7).","section":"§4.1, Table 5"},{"comment":"The central accuracy claim is supported only by visual profile comparisons in Figs 7 and 8; no quantitative error metric, uncertainty estimate, or convergence data is reported. This makes it impossible to judge the magnitude of the PMoE improvement over the SA baseline or to compare cases. Please add error norms (e.g., L2 relative errors in mean velocity and secondary-flow components) for trained and unseen cases.","section":"§4.2–4.3"},{"comment":"The information-bottleneck (IB) objective is presented as the training objective of the autoencoder, but the actual loss (Eq. 2.5) is reconstruction MSE and Eq. (2.8) is never optimized. The statement that a minimal feature subset S satisfying Eq. (2.9) follows from the IB Lagrangian is unsupported. This is a methodological overreach that should be removed or replaced by a derivation showing how reconstruction training implements IB.","section":"§2.2.2, Eqs. (2.8)–(2.9)"},{"comment":"The acceptance threshold T_accept=90% and the sampling threshold epsilon=0.05 are fixed and calibrated on the four regimes considered; no sensitivity analysis is given, and the paper admits T_accept would need tuning in future applications. The claim that the router 'detects when a flow is not represented' is therefore demonstrated only for well-separated regimes with the chosen features. A concrete test on a held-out flow family (e.g., a curved duct or airfoil separation) and a sensitivity study of T_accept would be needed to support the novelty-detection claim.","section":"§2.2.2 and §4"}],"minor_comments":[{"comment":"Typos: 'weel' should be 'well' (§3.2); 'shwon' should be 'shown' (§3.4); 'at at' duplicated (§4.2).","section":"§3.2, §3.4, §4.2"},{"comment":"Caption contains 'DNSbaselineexpert' with missing separators; format consistently throughout.","section":"Figure 5"},{"comment":"The PFI analysis is presented as motivation, but the router uses the full seven features. Clarify whether PFI is an interpretability diagnostic only or actually informs feature selection.","section":"§2.2.1, Appendix A"},{"comment":"The term 'confidence' for p_k is potentially misleading, as p_k is a recognition rate based on reconstruction-error thresholding, not a calibrated probability. Consider renaming it 'recognition rate' or similar.","section":"§2.2.2"},{"comment":"No data or code availability statement is provided. Please include access details for training data and expert implementations, as reproducibility is important for a data-driven modelling paper.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about the threshold inconsistency is valid and load-bearing: the 99.9th-percentile protocol contradicts the reported confidences in Tables 3 and 5. I would ask the authors to resolve this and to add quantitative error metrics before publication. The paper's novelty is adequate for JFM and the unseen-case evidence is promising, but the router validation needs more work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper deserves a serious look, but not a clean pass. The PMoE architecture—a modular bank of autoencoders routing to sparsely activated, heterogeneous RANS experts—is a genuine integration of known components into something I haven't seen before. The demonstration on four flow regimes is coherent, and the a posteriori results are directionally positive: the final PMoE-S3 model improves on baseline SA for the trained cases and for same-family extrapolations (different Reynolds numbers, different hill slopes) without disturbing older experts. The forgetting tests are reassuring, and the computational overhead is small. Those are real strengths.\n\nThe soft spots are proportionate but load-bearing. First, the router's novelty detection has an internal inconsistency. The paper states that T_k is set to the 99.9th percentile of the training data's reconstruction error (§2.2.2). By Eq. (2.6), the confidence on that same data should then be about 99.9%. Yet Table 3 reports the new component's confidence for separated flow as 93.5% and corner flow as 97.0%; Table 5 shows the training cases PH1p0 and SD5693 at 91.5% and 97.0%. Unless the threshold is not actually the 99.9th percentile, or the confidence is computed on a different sample distribution, the protocol is misreported. This matters because the entire continual-learning claim rests on the router reliably flagging a flow as unknown. The 'novel regime' detection is not yet validated as specified.\n\nSecond, the validation is qualitative: plots without quantitative error metrics, no repeated runs, no error bars. The 'unseen' cases are all within the same flow families; a genuinely new geometry or physics is not tested. Third, no code, data, or weights are released, which hampers independent verification.\n\nThe idea is worth referee time—the issues are fixable, not fundamental. But the router discrepancy needs a clear explanation, and the claims should be scaled to what is actually demonstrated. I would not cite this in my own work until a corrected version with code and quantitative metrics appears.","headline":"A promising modular continual-learning architecture for RANS, but the router's novelty-detection protocol is internally inconsistent and needs fixing before the core claim can be trusted.","tokens_in":23038,"tokens_out":5520,"would_cite":false,"duration_ms":52603,"reading_group":"yes","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Turbulence models can learn new flow regimes without retraining: an autoencoder router adds a specialist and leaves old ones untouched.","keywords":["continual learning","RANS turbulence modelling","mixture-of-experts","autoencoder routing","catastrophic forgetting","flow regime classification","data-driven turbulence closure","Reynolds-averaged Navier–Stokes"],"falsifier":"Feed the router a flow regime that lies between the trained ones—for example a boundary layer with a mild adverse pressure gradient that shares wall-attached features but develops a small separation bubble, or a periodic hill with a slope intermediate between 1.0 and 1.2—and record the winning expert's confidence and the resulting RANS prediction. If the router assigns more than 90% confidence to the wrong expert and the prediction degrades relative to the baseline Spalart–Allmaras model, the core claim of reliable novelty detection fails.","tokens_in":22093,"feed_emoji":"🌊","tokens_out":5807,"duration_ms":57217,"temperature":0.7,"pith_summary":"The paper claims that a turbulence model for engineering fluid flow can be built incrementally, one flow regime at a time, without ever retraining the whole model. The proposed progressive mixture-of-experts framework uses a bank of small autoencoders as a router: each autoencoder learns the signature of one flow type, and a flow is recognized when the fraction of sampled points whose reconstruction error falls below a threshold is high enough. When a flow is not recognized, the system trains one new autoencoder and one specialist correction, while all existing modules stay frozen, so older skills are preserved by construction. The claim is that this accumulation of specialists, demonstrated on airfoil wakes, channel flows, periodic-hill separation, and square-duct corner flows, predicts both seen and unseen cases better than the baseline Spalart–Allmaras model, at almost constant inference cost. A reader should care because this offers a concrete path toward industrial RANS models that grow with new data instead of being rebuilt from scratch.","feed_headline":"Mixture-of-experts learns new flow regimes without forgetting old ones","feed_subtitle":"A bank of autoencoders detects unfamiliar flows, then a dedicated correction is added while previous experts stay untouched.","key_machinery":"The load-bearing mechanism is the modular autoencoder router. Each router component has a small encoder that compresses the seven input features into a four-node bottleneck and a decoder that reconstructs them; the reconstruction error along the full encoder–decoder path, compared with a per-component threshold, becomes a confidence via the fraction of accepted points. A winner-takes-all rule selects the expert with the highest confidence, but activation occurs only if that confidence clears a global 90% acceptance threshold; otherwise a new component and a new expert are added. The experts themselves are heterogeneous—a symbolic damping-function correction for attached flows, a neural produ","core_discovery":"The paper's central claim is a continual-learning recipe for RANS closures: keep a family of specialist corrections and choose among them with a bank of autoencoders. Each autoencoder reconstructs seven local flow features of one regime; the fraction of sampled points with reconstruction error below the component's threshold gives a confidence, and the winning component is activated only if that confidence exceeds 90%. Below 90%, the flow is flagged as new, and one new autoencoder plus one new expert are trained while old modules stay untouched—catastrophic forgetting is avoided by design. Because only one expert is active at a time, adding experts costs almost nothing at inference. Validati","pith_inferences":["The paper leaves implicit that the 90% acceptance threshold and the fixed seven-feature input set are engineering choices; a natural extension is to calibrate the threshold per component or make it adaptive, since real industrial flows will rarely present four cleanly separated regimes.","A logical next test would be region-aware routing: the current router classifies an entire flow field by global statistics, but a single simulation can contain two regimes at once (for example, an attached boundary layer upstream of a separation bubble), and routing locally would be needed to handle that.","If the reconstruction signal really tracks the features that matter for each regime, the per-regime feature-importance profiles reported in the appendix double as an interpretability tool, potentially guiding which correction form a human engineer should choose for a new flow.","The method could be stress-tested first on deliberately ambiguous intermediate flows—cases that share wall-attached and separated features—to find where the novelty threshold starts to fail."],"forward_implications":["New flow regimes can be added to a deployed RANS model by training only one new module, leaving all prior experts untouched and preserving their predictions exactly.","Because only the winning expert is activated, the inference cost of the final model remains close to that of a single expert; the paper reports total wall-clock increases of 0.15–4% across test cases.","The same reconstruction-error signal provides an unsupervised way to flag out-of-distribution flows during a CFD run, not just to classify a fixed set of known regimes.","The framework can mix symbolic and neural corrections within one model, so future closure advances can be absorbed without changing the routing infrastructure.","Validation on unseen channel Reynolds numbers, square-duct Reynolds numbers, and three hill slopes suggests the learned corrections transfer beyond their exact training conditions."],"fun_headline_variants":["Autoencoder router adds new flow experts without forgetting old ones","Continual turbulence modeling: new regimes, no forgetting, same inference cost","Mixture-of-experts learns unseen flow regimes while keeping previous experts","Sparse experts: add turbulence corrections, keep prior specialists intact","Progressive MoE: autoencoder picks the right closure, expands without retraining"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The system stands or falls on the assumption that the reconstruction error of the seven chosen local features, compressed through a four-node bottleneck and compared with a fixed 90% acceptance threshold, reliably separates every future flow regime from the ones already seen; if two regimes look alike to the autoencoders or the threshold is miscalibrated, the router will pick the wrong expert or spawn redundant ones.","fun_headline_variants_meta":{"raw":{"variants":["Autoencoder router adds new flow experts without forgetting old ones","Continual turbulence modeling: new regimes, no forgetting, same inference cost","Mixture-of-experts learns unseen flow regimes while keeping previous experts","Sparse experts: add turbulence corrections, keep prior specialists intact","Progressive MoE: autoencoder picks the right closure, expands without retraining"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000217,"raw_usage":{"total_tokens":1262,"prompt_tokens":726,"completion_tokens":536,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":470,"completion_tokens_details":{"reasoning_tokens":444}},"tokens_in":470,"tokens_out":536,"duration_ms":5572,"temperature":1.0,"reasoning_tokens":444,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T10:37:36.589100+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Feed the router a flow regime that lies between the trained ones—for example a boundary layer with a mild adverse pressure gradient that shares wall-attached features but develops a small separation bubble, or a periodic hill with a slope intermediate between 1.0 and 1.2—and record the winning expert's confidence and the resulting RANS prediction. If the router assigns more than 90% confidence to the wrong expert and the prediction degrades relative to the baseline Spalart–Allmaras model, the core claim of reliable novelty detection fails.","supporting_citations":[],"review_version":1}