{"id":"d944a329-e318-4016-8814-c821b7b94d73","arxiv_id":"2501.04400","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A domain-decomposed coupling of Operator Inference and sparse full-order model inference that predicts localized transport dynamics beyond training data using a new Gershgorin-based stability regularization.","lead":"This paper combines two data-driven modeling techniques, one that works on a compressed low-dimensional space and one that works on the full simulation grid, to simulate dynamical systems whose interesting behavior is concentrated in a small region. The combined model predicts wave propagation and ice-thickness changes beyond the conditions used for training, with roughly 8x faster online simulations in a glacier test case.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Burgers 'beyond training' claim is not established for the full T=18 horizon: periodic wrap sends the front from the sFOM subdomain back into the OpInf subdomain, a regime the paper's Appendix A shows to be unstable.","rationale":"The reader's weakest assumption is the fixed a priori subdomain partition. My concern is the sharpest consequence of that assumption: in the periodic Burgers example, the localized feature does not stay inside the manually selected sFOM subdomain over the full stated prediction horizon, because periodicity sends it back into the OpInf subdomain. This makes the main 'beyond training' demonstration hinge on an unstated temporal limit or on undocumented periodic-interface coupling. The paper deserves credit for releasing code, for the honest Appendix A parameter study, and for the Gershgorin regularization, but the Burgers experiment is the only direct evidence for the central extrapolation claim, and its reported error is not decomposed in time around the wrap. A single run of the released code with a time-resolved error log settles whether the claim holds for the full 18s horizon. If the concern lands, the paper remains a useful CONDITIONAL contribution, but the claim should be qualified; it does not require rejection because the glacier experiment and the general methodology may still be valid for features that stay within the sFOM region.","tokens_in":24994,"tokens_out":11463,"duration_ms":123956,"concrete_test":"Using the released repository (github.com/lgkimisis/opinf_sfom), reproduce the Burgers experiment with the stated parameters and log the coupled model error in two windows, t in [9,10] and t in [10,18], and record the time at which the front's center crosses z=10/0. If the error is small only before the wrap and grows after the front enters the OpInf subdomain, the extrapolation claim is limited to a shorter horizon and the periodic interface is unmodeled. If the error remains small after wrap, inspect the code for explicit coupling at the z=0/10 periodic boundary and report it, because it is absent from Eq. (17) and Section 3.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central 'beyond training' demonstration in Section 5.1 uses periodic Burgers (Eq. 34) with c=0.5, L=10, T=18, training on t<=9, and the fixed partition z in [0,5) for OpInf, z in [5,10] for sFOM. The front travels toward z=L at speed ~c*w ~ O(0.5), so it reaches z=10 near t=10 and, by periodicity, re-enters z=0, which lies inside the OpInf subdomain, for roughly the second half of the stated horizon. The OpInf basis was chosen because the [0,5) training data have fast singular value decay; it is not designed to represent a transported front. Appendix A confirms this failure mode: if the interface is moved to a>5.5 so that transport-dominated dynamics enter the OpInf subdomain, the inferred model is unstable. The formulation in Sections 3.1.2-3.1.4 also describes only one OpInf-sFOM interface, but on a periodic domain with a single cut at z=5 there is a second interface at z=0/10; the paper does not state how periodic coupling across that interface is included in the inference problem (17). Thus the headline claim of accurate predictions beyond the span of the training snapshots is only supported for times before the front wraps, unless the released code contains additional periodic-interface handling that is not documented. The paper reports no error separately for t in [9,10] versus t in [10,18], so this window is not checked.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a domain-decomposition framework coupling Operator Inference (OpInf) in subdomains with fast singular value decay and sparse Full-Order Model (sFOM) inference in subdomains with slow decay, together with a Gershgorin-theorem-based regularization that promotes stability of the inferred linear operators. The coupled OpInf-sFOM method is applied to a 1D Burgers equation with periodic boundary conditions (training on t<=9 s, evaluation to T=18 s) and to a 2D parametric Pine Island Glacier ice-thickness model (training on 15 years at the extreme melt rates alpha=0 and alpha=100 m/yr, evaluation for intermediate alpha and up to 20 years). The paper claims accurate predictions beyond the span of the training snapshots, with an approximately 8x online speedup in the glacier test. A parametric cost analysis for offline and online phases is included, along with a released Python code for the Burgers example.","tokens_in":25473,"tokens_out":11429,"duration_ms":104766,"significance":"If the central claims are established, the work offers a practical way to extend non-intrusive model reduction to problems with spatially localized, transport-dominated features, and the Gershgorin regularization is a useful addition to the OpInf/sFOM toolkit. The Burgers code is publicly available, and the parametric cost formulas are a helpful planning tool. However, as detailed below, the evidence for the headline extrapolation claim is incomplete, and the treatment of the nonlinear threshold in the glacier test needs clarification.","major_comments":[{"comment":"The Burgers test with periodic boundary conditions and T=18 s does not establish the claimed 'beyond training' predictions for the full stated horizon. With the interface at z=5, the front, which travels at speed roughly c*w, exits through z=10 near t=10 s and re-enters the OpInf subdomain [0,5) via the periodic boundary. Appendix A shows that when transport-dominated dynamics enter the OpInf subdomain (interface a>=5.5), the inferred models become unstable. Since no error metrics are reported separately for t in [9,10] versus t in [10,18], the demonstration only supports extrapolation before the wrap-around. The authors should either restrict the claim to that interval, provide time-resolved errors showing the model remains accurate after wrap, or add an explicit periodic-interface treatment.","section":"Section 5.1, Appendix A"},{"comment":"The role of the nonlinear basal-melt threshold in the sFOM inference is underspecified. The sFOM formulation in Sections 2.3 and 3.1 assumes the quadratic structure of Eq. (2)/(18), but the indicator forcing m_b(h,alpha)=alpha*1_{h>h_f} in Eq. (37) is not polynomial. The text states that 'we infer ... a nonlinear sFOM model (due to (36)) by solving (16)', but Eq. (16) is the linear coupled system, not an inference problem, and it is not explained how the threshold is represented in the least-squares problem (11) or in the coupled equations (18). Without this explanation, the glacier test cannot be evaluated as a demonstration of the proposed OpInf-sFOM framework. Please clarify whether m_b is a known state-dependent forcing, an additional unknown operator, or a precomputed input, and how the non-smooth indicator is handled during inference and online integration.","section":"Section 5.2.2, Eq. (37)"},{"comment":"The abstract and conclusions claim 'an average prediction error on the order of 1%', but Figure 13 reports median errors that rise to around 1% only at later times, with boxplot maxima of order 10%. The exact metric (spatial average, time average, or median) is not defined in the text, which makes the headline claim difficult to reproduce. Please define the error metric precisely and report it separately for the OpInf and sFOM subdomains, as the spatial distribution of error is directly relevant to the domain-decomposition claim.","section":"Abstract and Section 5.2.2, Figure 13"},{"comment":"The fixed a priori subdomain partition is a limiting assumption that should be stated more prominently. The paper acknowledges in Section 6 that an adaptive decomposition is future work, but the abstract and introduction present the method as enabling predictions 'beyond those sampled in the training data set' without this caveat. Since Appendix A shows that a slightly larger OpInf subdomain leads to unstable models, the method's validity depends on the localized feature remaining inside the manually chosen sFOM subdomain for the entire prediction horizon. The authors should state this condition in the abstract and conclusions and discuss how the singular-value-decay indicator of Section 3.1.1 could be used to check it beyond the training window.","section":"Section 3.1.1, Appendix A, Section 6"}],"minor_comments":[{"comment":"The phrase 'by solving (16)' should be 'by solving (17)' or '(18)', since Eq. (16) is the coupled system, not the inference problem.","section":"Section 5.2.2"},{"comment":"The statement '(x(t) (x) x(t))_{E_i} = x_{Q_i}(t) (x) x_{Q_i}(t)' relies on an implicit ordering of the Kronecker product; please define the index set E_i accordingly so that this identity holds.","section":"Section 2.3"},{"comment":"The axis labels in the left and right panels are inconsistent with the text: the left panel's x-axis reads 'ROM dimension / global sFOM DOFs' while the right panel refers to 'ROM dimension / global ROM dimension'; please align the labels with the text and with Eqs. (30)-(31).","section":"Figure 5"},{"comment":"The sentence 'equation (2) with T and all higher order terms equal to zero' contains an undefined symbol T; presumably a typo for the quadratic term.","section":"Section 5.1"},{"comment":"Only the Burgers code is released; please state explicitly that the glacier data and code are not publicly available, and where possible provide the sFOM inference details needed to reproduce the glacier results.","section":"Data & Code Availability"},{"comment":"The cost analysis assumes n_T >> m; it would be helpful to state the actual values of n_T and m for the two test cases to verify that this assumption is satisfied.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely problem, and the released Burgers code is a strength. The main concern is the unverified post-wrap behavior in the Burgers test, which the authors' own appendix shows to be a failure mode; this needs to be either fixed or the claims carefully scoped. The glacier test's treatment of the discontinuous melt threshold is the other issue that must be resolved before the paper can be considered. If these are addressed, the paper could be a good fit for the journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hope you're well. Quick take on arXiv:2501.04400.\n\nThe real news here is coupling Operator Inference with sparse-FOM inference in a domain-decomposed, fully non-intrusive setup. Prior work coupled OpInf to OpInf or OpInf to a first-principles FOM, so the OpInf-sFOM combination is new, and it makes sense for systems with a spatially localized region of slow singular value decay. The Gershgorin-regularization trick, with a closed-form solution, is a useful addition that promotes stable linear operators without a constrained solve. The paper also ships Burgers code and data, which is good practice.\n\nThe glacier test is the most convincing part. It trains on the two extreme melt rates and 15 years, then predicts intermediate rates and up to 20 years. The 1% error and ~8x speedup are plausible, and the floating/grounded misclassification ~4.5% for α=0 is honestly reported. This is a genuine forward prediction, not curve-fitting.\n\nNow the soft spots, in proportion. First, the Burgers 'beyond training' claim is not established over the full T=18 horizon. The front travels at speed ~0.5, hits z=10 around t=10, and by periodicity re-enters z=0, which is inside the OpInf subdomain. The OpInf basis was trained on the [0,5) data from t<=9, where the dynamics are smooth; it never saw a transported front, and Appendix A shows that putting transport-dominated dynamics into the OpInf subdomain produces unstable models. The paper never reports a time-split error, so we don't know how the model behaves after wrap. Maybe the released code handles the periodic second interface, but the text doesn't say how. This is a fixable presentation issue: either report the error separately for t in [9,10] and [10,18], or qualify the claim.\n\nSecond, the subdomain split is chosen a priori and fixed; Appendix A demonstrates the fragility. The authors acknowledge this, so it's a stated limitation rather than a hidden one, but it does narrow the scope to problems where the localized feature stays inside the sFOM subdomain during prediction.\n\nThird, the glacier model's nonlinearity is underspecified. Section 5.2.2 says a linear OpInf and a nonlinear sFOM are inferred 'by solving (16)', which is a linear system. The basal-melt Heaviside (Eq. 37) is not a polynomial term, and the paper doesn't explain how it enters the sFOM inference or how it is evaluated during prediction. That needs a clear paragraph for the method to be reproducible.\n\nWho is this for? Researchers working on non-intrusive ROM for transport-dominated or localized-feature dynamics. They'll get a clear formulation, a sensible cost analysis, and a useful application. It deserves a proper peer review; with the Burgers horizon point addressed and the ice-model nonlinearity clarified, it's a solid publication. I'd send it to review.","headline":"OpInf-sFOM domain-decomposed non-intrusive ROM is a useful new combination, but the Burgers 'beyond training' claim doesn't cover the full periodic horizon, and the glacier nonlinearity is underspecified.","tokens_in":25963,"tokens_out":6422,"would_cite":true,"duration_ms":59948,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["37M99","65M99"],"pacs":[],"model":"deepseek-v4-flash","headline":"A domain split lets reduced models predict beyond the training data.","keywords":["Operator Inference","sparse full-order model inference","domain decomposition","transport-dominated dynamics","singular value decay","Gershgorin regularization","non-intrusive reduced-order modeling","ice thickness prediction"],"falsifier":"Run the coupled OpInf-sFOM on a transport-dominated test case whose localized feature crosses the prescribed interface during the prediction period; the model should show accuracy collapse or instability (as the paper's Appendix A already does for a Burgers interface beyond a=5.5), confirming the fixed partition is the decisive limit.","tokens_in":21,"feed_emoji":"🧊","tokens_out":4772,"duration_ms":108710,"temperature":0.7,"pith_summary":"This paper argues that the obstacle to reduced-order modeling of transport-dominated systems—slowly decaying singular values—can be bypassed by splitting the spatial domain before learning. On the part where dynamics are low-dimensional, it infers an Operator Inference (OpInf) reduced model; on the part where a localized feature such as a moving front or melt region lives, it infers a sparse full-order model (sFOM). The two are coupled through interface inputs, with a post-processing interpolation for smoothness. The paper shows this coupled OpInf-sFOM predicts Burgers shocks beyond the span of training snapshots, and for a Pine Island Glacier ice-thickness model achieves an average error near 1% with an online speedup of roughly 8x. If true, this gives non-intrusive models a way past the training-snapshot-span limitation for spatially localized slow singular value decay.","feed_headline":"Domain split lets reduced models predict beyond training data","feed_subtitle":"Sparse full-order inference handles the moving front while Operator Inference compresses the rest, yielding ~1% glacier errors at ~8x speed.","key_machinery":"The central mechanism is the domain decomposition of the state vector into an OpInf subdomain and an sFOM subdomain, selected by comparing normalized singular value decays of subdomain snapshot matrices. OpInf learns a low-dimensional quadratic (or linear) model from projected snapshots, while sFOM learns adjacency-based stencil coefficients at full order for each DOF; coupling terms across the interface enter each subproblem as inputs. To promote stable inferred operators, the paper adds a Gershgorin-disk-based regularizer that penalizes the L2 norm of operator rows (bounding disk radii) and adds a diagonal penalty that pushes disk centers toward the left half-plane, both with closed-form least-squares solutions. A post-processing interpolation over an overlapping strip enforces solution smoothness across the interface.","core_discovery":"The central claim is that coupling a projection-based ROM to a full-order inferred model through domain decomposition yields non-intrusive predictions that are not confined to the span of the training snapshots. In the proposed formulation, the domain is partitioned a priori based on the gap in singular value decay: the subdomain with fast decay is reduced through Operator Inference, while the subdomain exhibiting slow decay is modeled by sparse full-order inference, which is not bound to a reduced basis. The two subproblems are coupled by interface DOFs, and a Gershgorin-disk-based regularization with closed-form solution drives inferred linear operators toward stability. The paper demonstrates the coupled model on a 1D Burgers' equation with a propagating wave beyond the training interval and on a parametric 2D ice-thickness model of Pine Island Glacier, reporting average prediction errors on the order of 1% and an average online speedup of 7.89x, roughly 8x, over the full-order code.","pith_inferences":["An adaptive interface that moves the sFOM subdomain to track a propagating feature would extend the method to problems where the localized region travels; the paper's appendix shows that a fixed interface fails once transport-dominated dynamics enter the OpInf subdomain.","The singular-value-decay gap indicator could be combined with an online error estimator to trigger subdomain re-partitioning, opening a path to reliable a posteriori control of the decomposition.","The same Gershgorin penalty idea applies to any least-squares-based non-intrusive model with a linear term, so it may improve robustness in other inference pipelines beyond OpInf-sFOM.","Because full-order inference can represent local dynamics that a basis cannot, the method may be suited to predicting local bifurcations, although the paper only lists this as future work."],"forward_implications":["Accurate non-intrusive predictions are possible for transport-dominated features that evolve beyond the span of the training snapshots, as shown by the Burgers' wave test.","A parametric ROM can be trained on extreme parameter values and predict intermediate cases, as demonstrated by the ice-thickness model trained only at alpha=0 and alpha=100 m/yr.","The online cost savings grow as the slow-decay region shrinks relative to the full domain and as the OpInf dimension decreases; the paper gives explicit cost ratios for estimating this.","Stable inferred linear operators are promoted by the proposed Gershgorin regularization, which the paper shows produces negative-real-part eigenvalues for both OpInf and sFOM operators in the test cases."],"supporting_citations":[{"why":"Supplies the Operator Inference method used on the low-dimensional subdomain.","marker":"[47]"},{"why":"Supplies the sparse full-order model inference method and its property of predicting beyond the training snapshot span.","marker":"[62]"},{"why":"Provides the original ROM-FOM domain decomposition idea for shock-containing flows, which this paper adapts to a non-intrusive setting.","marker":"[38]"},{"why":"Demonstrates domain-decomposed OpInf for large-scale systems and the interface interpolation step used here.","marker":"[16]"},{"why":"Offers a constrained-regression approach for stable sparse differential operators, which motivates the stability-promoting regularization.","marker":"[48]"},{"why":"Provides physics-informed regularization for learning stable OpInf models, a key precursor to the Gershgorin regularizer.","marker":"[55]"},{"why":"Introduces stencil inference for discrete differential operators, the conceptual basis of sFOM.","marker":"[61]"},{"why":"Supports the adjacency-based sparsity pattern used for sFOM inference in fluid-structure applications.","marker":"[21]"}],"fun_headline_variants":["Coupled ROMs predict beyond training data via domain split","Hybrid ROM tackles localized features with 8x speedup","Domain-decomposed modeling beats extrapolation limit","OpInf + sFOM: non-intrusive ROM for localized dynamics","Split-domain ROM: predict outside trained range, ~1% error"],"cache_read_input_tokens":28032,"weakest_assumption_plain":"The load-bearing premise is that the subdomain partition is fixed before training: the moving front or localized feature must remain inside the chosen sFOM subdomain over the whole prediction horizon, since there is no adaptive mechanism to move the interface.","fun_headline_variants_meta":{"raw":{"variants":["Coupled ROMs predict beyond training data via domain split","Hybrid ROM tackles localized features with 8x speedup","Domain-decomposed modeling beats extrapolation limit","OpInf + sFOM: non-intrusive ROM for localized dynamics","Split-domain ROM: predict outside trained range, ~1% error"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000643,"raw_usage":{"total_tokens":3013,"prompt_tokens":1054,"completion_tokens":1959,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":670,"completion_tokens_details":{"reasoning_tokens":1873}},"tokens_in":670,"tokens_out":1959,"duration_ms":12191,"temperature":1.0,"reasoning_tokens":1873,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:34:14.315691+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the coupled OpInf-sFOM on a transport-dominated test case whose localized feature crosses the prescribed interface during the prediction period; the model should show accuracy collapse or instability (as the paper's Appendix A already does for a Burgers interface beyond a=5.5), confirming the fixed partition is the decisive limit.","supporting_citations":[{"cited_title":"Schumann and P","cited_arxiv_id":null,"evidence_quote":"Supplies the sparse full-order model inference method and its property of predicting beyond the training snapshot span."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the original ROM-FOM domain decomposition idea for shock-containing flows, which this paper adapts to a non-intrusive setting."},{"cited_title":"Prakash and Y","cited_arxiv_id":null,"evidence_quote":"Offers a constrained-regression approach for stable sparse differential operators, which motivates the stability-promoting regularization."}],"review_version":1}