{"id":"ca48cb54-f369-4cf4-a3e5-b28f45109418","arxiv_id":"2502.12164","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A physics-informed graph neural network with two-phase training and physics-preserving normalization estimates hydraulic states in water networks, improving scalability and out-of-distribution robustness over the prior PI-GNN.","lead":"This paper introduces a graph neural network that estimates water pressure and flow in pipe networks, trained with hydraulic equations and a new normalization scheme. It reports faster and more robust predictions than the previous model, especially on larger networks with pumps and pressure-reducing valves.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Normalization premise (largest flow ≈ total demand) fails for multi-reservoir and pumped WDSs; gradient-control claim is unsupported.","rationale":"The reader's weakest assumption correctly identifies the demand-sum normalization as load-bearing, but frames the failure as biased de-normalization. I think the more precise failure is the gradient-control argument: normalized flows can exceed 1 in multi-reservoir or pumped networks, so the claimed mechanism for avoiding exploding gradients is not in force, while the reversibility algebra itself remains exact. This concern does not by itself overturn the paper: the empirical results could still hold, enabled instead by two-phase training and gradient clipping. It does, however, weaken the stated justification for a central methodological contribution and should be tested before accepting the scalability claim at face value. The reader's verdict of CONDITIONAL already captures the need for such validation, so I would leave the verdict unchanged.","tokens_in":12863,"tokens_out":12579,"duration_ms":132877,"concrete_test":"Run EPANET/WNTR on all eleven WDSs over the same sampled demand and diameter ranges and compute R = max_e |q_e| / Σ_v d_v per sample. If the distribution of R stays at or below 1 across all samples, the premise survives and this objection fails. If R exceeds 1 substantially on any multi-reservoir or pumped WDS, retrain the affected networks with the demand-sum normalization replaced by a normalization based on the actual maximum flow and compare gradient norms and Table III; this isolates whether the demand-sum normalization is load-bearing for the scalability claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The scalability/robustness argument in Sections III-D and IV-B assumes that dividing demands by their sum keeps the largest flow 'around 1' because the largest flow in a WDS is close to the sum of all demands. This premise does not hold for the benchmark suite. In a network with two reservoirs at different heads, a pipe directly connecting them carries q = ((Δh)/r)^(1/1.852), which is independent of consumer demand; consumer demand can be made arbitrarily small, so max|q|/Σd is unbounded. Pumps can similarly drive recirculating flow that exceeds total demand. The paper evaluates several multi-reservoir and pumped WDSs (Anytown, Pescara, Modena, PA1, Balerma, L-Town Area A, L-Town All Areas), so in those cases normalized flows can be much larger than 1. The algebra of the normalization is still reversible, so the failure is not 'biased de-normalization'; the load-bearing casualty is the stated mechanism for eradicating exploding gradients and thereby enabling stable training and scalability. Because gradient clipping is also used (Section IV-B), the distinct contribution of the demand-sum normalization to the large-network and OOD results is not established without additional evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SPI-GNN, a physics-informed graph neural network surrogate for hydraulic state estimation in water distribution systems (WDSs). It extends a prior PI-GNN model by adding pumps and pressure-reducing valves, a two-phase training scheme, a flow-aware message-passing architecture, and a physics-preserving data normalization method. The authors evaluate on eleven WDSs of increasing size, reporting in-distribution accuracy against EPANET/WNTR, robustness to out-of-distribution demands and pipe diameters, and large speed-ups over the hydraulic simulator.","tokens_in":13101,"tokens_out":5457,"duration_ms":54397,"significance":"If the empirical results hold, the paper would deliver a practically relevant surrogate model that is substantially faster than EPANET and more scalable than the prior PI-GNN baseline, with a comparatively large benchmark suite of eleven WDSs and a systematic out-of-distribution protocol. The evaluation against an independent simulator (EPANET/WNTR) and the use of a fixed, non-learned physical correction step are genuine strengths. However, the central justification for the proposed normalization mechanism rests on an unverified premise about maximum flows, and several claims exceed the evidence presented, especially the blanket statement of state-of-the-art superiority and the claim of greater out-of-distribution robustness without a direct baseline comparison.","major_comments":[{"comment":"The central justification for the demand-sum normalization is the statement in Section III-D that 'the largest flow in a WDS is close to the sum of all demands in the WDS', and Section IV-B repeats that this restricts the largest flow to around 1 and eradicates exploding gradients. This premise fails for several networks in Table I that have multiple reservoirs or pumps (Anytown, Pescara, Modena, PA1, Balerma, L-Town All Areas). In a pipe connecting two reservoirs at different heads, the flow is driven by the head difference and is independent of consumer demand; pumps can similarly drive recirculating flows that exceed total demand. The normalization algebra is reversible, so de-normalization is not biased, but the claimed gradient-control mechanism is not established. Moreover, because gradient norm clipping is also used in Section IV-B, the paper does not isolate the contribution of the demand-sum normalization to training stability and scalability. Please replace the premise with a provable bound or an empirical gradient-norm analysis, or run an ablation that isolates the normalization from gradient clipping and the two-phase scheme.","section":"Sections III-D and IV-B"},{"comment":"The abstract's blanket claim that the model 'outperforms the current state-of-the-art DL model' is contradicted by the small-network results in Table III. On Hanoi, SPI-GNN has a higher head MAE than PI-GNN (0.013 vs 0.004); on Pescara, heads are 0.052 vs 0.049; on L-Town Area C, demands are 0.126 vs 0.117. The text in Section IV-D1 appropriately says SPI-GNN achieves 'similar MAE' for small WDSs, but the abstract and conclusion do not carry this qualification. The performance claim should either be restricted to larger networks or supported by a consistent superiority measure across all WDSs.","section":"Abstract and Table III"},{"comment":"The out-of-distribution experiments report only SPI-GNN's absolute mean MAE for demands and diameters, with no baseline comparison. The abstract claims the approach is 'more robust to out-of-distribution input features' than the state of the art, but Figures 2 and 3 contain no PI-GNN (or other surrogate) curves under the same protocol. These figures support absolute robustness of SPI-GNN on the tested WDSs, but they do not by themselves support the comparative robustness claim. Please add the baseline under identical OOD conditions, or rephrase the claim as absolute robustness rather than superiority.","section":"Section IV-D2 and IV-D3, Fig. 2 and Fig. 3"},{"comment":"The global physics-informed algorithm f2 is presented as an iterative scheme whose number of iterations J 'varies based on how close q_hat are to q*', but in the experiments K is fixed per WDS (Table II). The paper provides no convergence criterion, error bound, or empirical verification that the fixed iteration range yields a solution of Equations (2), (3), (5), and (7). Because this iterative algorithm is load-bearing for the physical consistency of the final predictions, please specify the stopping rule actually used and provide evidence that the fixed K is sufficient for each WDS, or report the sensitivity of the results to K.","section":"Section III-A2 and Table II"}],"minor_comments":[{"comment":"The notation in 'ˆqe := ¯qein ∥ (−¯qein)' mixes concatenation with scalar negation; please clarify the dimensions of the edge flow representation and how the two directions are assembled.","section":"Section III-A1"},{"comment":"The pump message equation contains typographical errors, including an unmatched parenthesis and an unclear combination of subscripts; please rewrite the formula cleanly and define all variables.","section":"Section III-A2"},{"comment":"The scaling factor τ = 1000 in the normalization of pipe resistance is introduced without justification; please explain how it is chosen and whether it affects the results.","section":"Section III-D"},{"comment":"The data-generation description says diameters are varied 'every 6th sample' while demands are varied across all samples; please clarify the exact sampling procedure for the training, validation, and test splits.","section":"Section IV-A"},{"comment":"The 'Demands (Estimated vs True)' column compares an input quantity (demands) with a model estimate; please explain what this metric measures and why it is informative given that demands are used as inputs.","section":"Table III"},{"comment":"The gradient clipping threshold is set to 0.00001, which is unusually small; please comment on how this threshold was selected and whether the final results are sensitive to it.","section":"Section IV-C"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the main risk is not circularity—the evaluation against EPANET/WNTR is sound—but the unverified normalization premise and the missing direct baseline in the out-of-distribution experiments. These are fixable with revision, but they affect load-bearing claims in the abstract and conclusion."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Know this before you read: the paper is an incremental but real step forward for GNN surrogates of water distribution systems. It inherits the PI-GNN architecture [8] and adds four things: bidirectional flow embeddings, a two-phase training scheme, support for pumps and pressure-reducing valves, and a demand-sum normalization. On the larger networks (Modena, PA1, Balerma, L-Town) it beats its predecessor by large margins, sometimes an order of magnitude in head error, and the robustness experiments sweep demand and diameter noise in a principled way. That part is worth taking seriously.\n\nThe soft spots are in the claims, not the engineering. The abstract says the model 'outperforms the current state-of-the-art DL model,' but Table III shows the opposite on small networks: Hanoi head MAE is 0.013 vs 0.004 for PI-GNN, and L-Town Area C demand MAE is 0.126 vs 0.117. The comparison also omits the edge-based GNN surrogates cited in Section II, so 'state of the art' is a narrow claim. No code or data is released, which limits how much a referee can check.\n\nThe bigger conceptual issue is the normalization. The paper justifies dividing demands by their sum with the claim that the largest flow in a WDS is close to the sum of all demands (Sections III-D and IV-B). That premise is false for multi-reservoir and pumped networks: a pipe between two reservoirs carries flow set by head difference, independent of consumer demand, and pumps can circulate flow larger than total demand. The stress-test note I saw makes this point correctly. The normalization algebra is still reversible, so de-normalized outputs are not biased; what collapses is the stated mechanism for 'eradicating exploding gradients.' Since the paper also uses two-phase training and gradient norm clipping, the unique contribution of the demand-sum normalization to the scalability results is not identified. A clean ablation would settle it.\n\nNone of this kills the paper. The empirical gains on large networks are large and the OOD evaluation is more thorough than most in the area. The citation pattern is normal for a continuation line; heavy self-citation to [8] is expected when you claim to improve on it. With code and a tempered abstract, this would be a solid contribution.\n\nI'd send it to review. A good referee will ask for an ablation of the normalization and a comparison to [33] and [34], but the work deserves that time.","headline":"Incremental but genuine extension of PI-GNN that scales to larger WDSs; the headline claims overstate the evidence and the normalization rationale is shaky, but the empirical work deserves review.","tokens_in":13654,"tokens_out":2337,"would_cite":false,"duration_ms":20958,"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":"The paper claims that a physics-informed graph neural network named SPI-GNN can estimate water pressure and flow in city-scale water distribution systems accurately enough to outperform the previous deep-learning surrogate, while scaling…","keywords":["water distribution systems","hydraulic state estimation","physics-informed graph neural networks","surrogate models","out-of-distribution robustness","data normalization","pumps and pressure-reducing valves","EPANET"],"falsifier":"Compute the ratio of the maximum pipe flow to the sum of demands in a benchmark water distribution network with pumps and loops; if this ratio exceeds 1 by a wide margin in realistic cases, the normalization premise is violated. A direct test: train SPI-GNN on such a network and compare de-normalized heads and flows against the hydraulic simulator on out-of-distribution demands—if the normalized MAE jumps well above the paper's reported levels, the scalability and robustness claims fail for recirculating systems.","tokens_in":12582,"feed_emoji":"💧","tokens_out":8411,"duration_ms":63793,"temperature":0.7,"pith_summary":"The paper is trying to establish that a deep-learning surrogate can replace slow hydraulic simulators for estimating pressure heads and pipe flows in water distribution networks, including realistic networks with pumps and pressure-reducing valves. The proposed model, SPI-GNN, couples a local graph neural network with a global physics-informed algorithm in a two-phase training loop, and uses a normalization scheme that respects the hydraulic equations. If the claims hold, engineers doing planning, expansion, and rehabilitation of water systems could run thousands of what-if simulations in a fraction of the time, on networks larger than previous surrogates could handle, without retraining for every new demand pattern. The paper reports accuracy comparable to or better than the prior PI-GNN model on eleven networks, with stable errors when test demands and pipe diameters move well outside training ranges.","feed_headline":"Physics-informed GNN predicts water flow and pressure across cities","feed_subtitle":"Beats the prior model on large networks; stays accurate when demands and pipe diameters shift far off-distribution.","key_machinery":"The load-bearing mechanism is the two-part model: a local learnable GNN ($f_1$) that estimates flows and demands by message passing, and a global physics-informed algorithm ($f_2$) that iteratively reconstructs pressure heads from pipe, pump, and PRV equations and then recomputes flows and demands from conservation of mass. The physics-preserving normalization is the other essential piece: demands are divided by their sum and pipe resistances by $3\\sigma_{r'}$, with reservoir heads, PRV settings, and pump parameters rescaled so that the normalized system obeys the same hydraulic laws and de-normalization is exact. Training alternates $T$ iterations of $f_1$ alone with $K-T$ iterations of $f_1$ plus $f_2$, and the loss ties estimated demands, physics-derived demands, and the two flow estimates together.","core_discovery":"The central claim is that a physics-informed GNN can perform hydraulic state estimation—recovering nodal pressure heads and link flows from demands, reservoir heads, pipe attributes, and pump and valve settings—accurately and efficiently on water distribution networks of practical size. Three modifications carry the argument: an architecture that replaces SeLU with ReLU and concatenates both directional flow embeddings; a training scheme split into a local GNN phase followed by a global physics-constrained phase that incorporates pumps and pressure-reducing valves; and a normalization that divides demands by their total and pipe resistances by three times their standard deviation, then rescales reservoir heads, PRV settings, and pump parameters so that the hydraulic equations stay intact. On eleven networks ranging from 19 to 935 junctions, the reported mean absolute errors on heads, flows, and demands are below a few percent, the model outperforms the previous state of the art on larger networks, and the error remains below 0.05 normalized MAE even when demand noise rises tenfold.","pith_inferences":["The normalization premise that the largest flow is close to the sum of all demands is plausible for tree-like gravity-fed networks but may fail in looped or heavily pumped systems where recirculating flows exceed total demand; testing on such a network would reveal whether the out-of-distribution robustness claim generalizes.","The reported out-of-distribution experiments vary demands and pipe diameters but not roughness coefficients, pump curves, or network topology; the normalization argument directly covers only the features it rescales.","A temporal extension with tank levels as state variables is needed before the surrogate can handle dynamic simulations, which the paper itself lists as future work.","The speed-up figures compare inference against a Python-based hydraulic solver; a comparison against optimized solver backends would give a more conservative efficiency estimate."],"forward_implications":["Engineers can use the surrogate in place of the standard hydraulic simulator for planning, expansion, and rehabilitation studies, cutting the cost of what-if simulations by orders of magnitude; the paper reports speed-ups from roughly 190x to over 21,000x depending on network and demand extremity.","The model extends surrogate state estimation to realistic networks with pumps and pressure-reducing valves, which the prior PI-GNN could not handle, and scales to networks of roughly one thousand junctions.","Because the surrogate is fully differentiable, it can be repurposed for gradient-based optimization tasks such as optimal sensor placement, pipe rehabilitation decisions, and estimating pressures everywhere from sparse sensor readings.","Out-of-distribution robustness means the surrogate can be applied to demand patterns and pipe diameter scenarios not present in training, as long as the normalization premise holds.","The two-phase training with a bounded iteration budget avoids the exploding gradients that prevented the prior model from training on larger networks."],"supporting_citations":[{"why":"The prior physics-informed GNN model that this paper extends and the main baseline it claims to outperform.","marker":"[8]"},{"why":"Supplies the hydraulic equations (conservation of mass and energy, pump and PRV relations) and the simulator used as ground truth.","marker":"[7]"},{"why":"Provides the WDS configuration files used to build the training and evaluation datasets.","marker":"[36]"},{"why":"Provides additional WDS configuration files for the larger and more realistic networks.","marker":"[37]"},{"why":"Used to run the hydraulic simulations that generate ground-truth heads, flows, and demands for comparison.","marker":"[38]"}],"fun_headline_variants":["Physics-informed GNN scales hydraulics to city networks","GNN hydraulic model stays accurate under 10x demand noise","Physics-aware GNN outperforms state-of-the-art on water grids","Surrogate GNN robust to shifts in demand and pipe size","Hydraulic GNN scales to large networks, beats prior model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method rests on the claim that the largest flow in a water distribution system is close to the sum of all consumer demands, so dividing demands by their sum keeps all flows near one; if looping or pumping creates recirculating flows much larger than total demand, the normalization no longer preserves hydraulics and the model's accuracy and out-of-distribution robustness would degrade.","fun_headline_variants_meta":{"raw":{"variants":["Physics-informed GNN scales hydraulics to city networks","GNN hydraulic model stays accurate under 10x demand noise","Physics-aware GNN outperforms state-of-the-art on water grids","Surrogate GNN robust to shifts in demand and pipe size","Hydraulic GNN scales to large networks, beats prior model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000999,"raw_usage":{"total_tokens":4204,"prompt_tokens":896,"completion_tokens":3308,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":3222}},"tokens_in":512,"tokens_out":3308,"duration_ms":22194,"temperature":1.0,"reasoning_tokens":3222,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T12:21:35.845126+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the ratio of the maximum pipe flow to the sum of demands in a benchmark water distribution network with pumps and loops; if this ratio exceeds 1 by a wide margin in realistic cases, the normalization premise is violated. A direct test: train SPI-GNN on such a network and compare de-normalized heads and flows against the hydraulic simulator on out-of-distribution demands—if the normalized MAE jumps well above the paper's reported levels, the scalability and robustness claims fail for recirculating systems.","supporting_citations":[{"cited_title":"Physics-informed graph neural networks for water distribution systems,","cited_arxiv_id":null,"evidence_quote":"The prior physics-informed GNN model that this paper extends and the main baseline it claims to outperform."},{"cited_title":"Epanet 2.2 user’s manual, water infrastructure division,","cited_arxiv_id":null,"evidence_quote":"Supplies the hydraulic equations (conservation of mass and energy, pump and PRV relations) and the simulator used as ground truth."},{"cited_title":"Epanet- benchmarks","cited_arxiv_id":null,"evidence_quote":"Provides the WDS configuration files used to build the training and evaluation datasets."},{"cited_title":"Waterbenchmarkhub,","cited_arxiv_id":null,"evidence_quote":"Provides additional WDS configuration files for the larger and more realistic networks."},{"cited_title":"An overview of the water network tool for resilience,","cited_arxiv_id":null,"evidence_quote":"Used to run the hydraulic simulations that generate ground-truth heads, flows, and demands for comparison."}],"review_version":1}