{"id":"3ecb93b7-e382-45fd-821f-ca42fdf0d810","arxiv_id":"1908.07794","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A calibration scheme based on explicit Colebrook-White turbulent flow plus a modified Newton-Raphson solver reconstructs pipe roughnesses and unmeasured pressures from noise-free head and consumption data.","lead":"This paper shows how to recover individual pipe roughness values in a water network by inverting the steady-state hydraulic equations, using pressure measurements at a subset of nodes during multiple loading conditions. A Newton-type solver with backtracking and multi-start is proposed, and a rule linking sensor count, pipe count, and measurement sets is derived.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Convergence proof for Algorithm 1 is invalid: Eq. (26) treats the rectangular Jacobian as invertible, so the L1 line search is not shown to be a descent direction.","rationale":"The reader's weakest assumption was the unverified full-rank condition of Assumption 5. That is a legitimate identifiability concern, but it is explicitly an assumption, and the paper's conditional statement acknowledges it. I focus instead on an internal correctness gap in the convergence argument: Eq. (26) and Algorithm 1's line search depend on an incorrect expression for the directional derivative of the L1 merit function. Even if Assumption 5 holds and J has full column rank, the Newton direction need not be a descent direction for the L1 norm, so the 'reliably' conclusion is not supported by the presented proof. This does not refute the numerical example, which may still have converged by good initialization and random restarts, but it strengthens the case for treating the paper as a feasibility demonstration rather than an established reliable method. The reader's CONDITIONAL verdict remains appropriate: the formulation and example are valuable, but either the descent proof must be corrected or much more extensive numerical evidence (including rank checks, noise, and varied networks) is needed before the reliability claim can be accepted.","tokens_in":18894,"tokens_out":14355,"duration_ms":149294,"concrete_test":"For the Section 6 example at the initial iterate x0 given in Eq. (36) (or any stored iterate), compute f0 = f(x0), J0 = J(x0), d = -(J0^T J0)^-1 J0^T f0, and the true directional derivative D = sum_i sign(f0_i) * [J0 d]_i. Compare D with s_k = -sign(f0)^T f0 as used in Algorithm 1 line 11. If D differs from s_k, or if D is positive, Eq. (26)'s descent claim fails for the paper's own problem. A minimal linear counterexample, e.g. J = [-1, -10, -10]^T with f = [-100, 1, 1]^T, verifies the algebraic error independently of the network application.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The claim that the algorithms 'enable to find the real root reliably' (Sec. 7) depends on Algorithm 1's convergence, whose supporting calculation, Eq. (26), is incorrect for this overdetermined problem. The Jacobian J in (20) is nm*nj by n_l + nm(nj-np), so it is not square; the Newton step uses the left inverse J^+ = (J^T J)^-1 J^T (Sec. 5). The directional derivative of v = ||f||_L1 along d = -J^+ f is -sign(f)^T J J^+ f, not -sign(f)^T f as written in Eq. (26). Since J J^+ is the orthogonal projection onto the column space of J, the identity J J^-1 f = f holds only when f lies in that column space, which is not generally true at a nonzero iterate of an overdetermined system. Indeed, even for a rank-one J one can choose f with a nonzero component orthogonal to range(J) so that sign(f)^T J J^+ f is negative, making the true directional derivative positive. Algorithm 1 line 11 and the line-search criterion (28) use g'(0) = -sign(f)^T f, which is not the actual slope of g(mu) = v(x + mu*d). Thus the Newton direction is not proved to be a descent direction, and the backtracking strategy rests on a false premise. The single noise-free example shows the method can work, but it does not establish the claimed reliability until the descent/line-search step is repaired or replaced.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an explicit inversion scheme for identifying per-pipe roughness coefficients in water distribution networks under fully turbulent flow. The authors formulate the steady-state hydraulic equations with Darcy-Weisbach friction and the implicit Colebrook-White relation rewritten in explicit turbulent flow form, obtaining a system (16) whose unknowns are the per-pipe roughnesses and the unmeasured nodal pressure heads for each measurement set. They derive the minimum number of independent measurement sets as nm = ceil(n_l / n_p), propose a modified Newton-Raphson algorithm with an L1 line search and a multi-start outer loop (Algorithms 1 and 2), and test it on a synthetic three-cycle network with eight pipes, three pressure sensors, and three measurement sets, reporting recovery of all roughnesses with a maximal deviation of about 6%.","tokens_in":19223,"tokens_out":5061,"duration_ms":50847,"significance":"The problem formulation is a useful addition to explicit calibration methods: it avoids the separate simulation-optimization loop typical of implicit calibration, makes the role of the number of measurement sets explicit, and the Jacobian structure in Eq. (20) is clearly laid out. The counting rule for independent measurement sets is simple and algorithmically relevant. However, the central reliability claim in Section 7 is not supported to the stated level. The convergence argument for the algorithm contains a mathematical error, identifiability is assumed rather than proven, and the numerical evidence is a single noise-free synthetic example. The contribution is therefore promising but requires substantial strengthening before the claimed reliability can be accepted.","major_comments":[{"comment":"The descent-direction calculation in Eq. (26) is invalid for the rectangular system used here. With the left-inverse Newton direction delta_x = -J^+ f, where J^+ = (J^T J)^{-1} J^T and J has size n_m n_j by (n_l + n_m(n_j - n_p)), the directional derivative of v = ||f||_{L1} is -sign(f)^T J J^+ f, not -sign(f)^T f. Since J J^+ is the orthogonal projection onto the range of J, the equality used in Eq. (26), and consequently g'(0) in Eq. (29) and the acceptance criterion (28), hold only when f lies in the column space of J, which is not generally true at a non-root iterate of an overdetermined system. The line-search control in Algorithm 1 therefore rests on an unverified premise, and the claim in Section 7 that the algorithms 'enable to find the real root reliably' is not supported.","section":"Section 5.2, Eq. (26)"},{"comment":"The full-column-rank condition on the stacked Jacobian in Eq. (20) is asserted, not proved. The counting inequality in Eq. (17) is necessary but not sufficient: it does not guarantee that a given network topology, sensor placement, and set of load variations produce a linearly independent stacked Jacobian. The paper does not verify the rank condition for the three-cycle example and does not characterize network or measurement conditions under which Assumption 5 holds. Because this assumption is the only mechanism separating pipe roughnesses from unmeasured nodal pressure heads, the identifiability of the problem is not established; if the rank condition fails, Eq. (16) is underdetermined and the algorithm cannot distinguish roughness from pressure unknowns.","section":"Assumption 5 and Eq. (20)"},{"comment":"The numerical evaluation is limited to a single noise-free synthetic scenario, yet Assumption 7 concerns measurement noise and the conclusion claims reliability under independent measurement sets. No noisy case is presented, and the authors themselves note in Section 6.3 that with measurement noise 'one certainly loses the capability to identify the real root by only looking at v(x)'. Furthermore, the simulated measurements are generated using the authors' own dynamic model from [15,16], whose equilibrium is equivalent to the steady-state equations (2) used for inversion; this is a self-consistency check rather than an independent validation. A Monte Carlo study over noise realizations, initial values, and network configurations, together with rank verification for each configuration, would be needed to substantiate the reliability claim.","section":"Section 6, Tables 3-4"}],"minor_comments":[{"comment":"The derivation of nm,min = ceil(n_l / n_p) is compressed; the underlying break-even condition m n_p >= n_l is not stated explicitly, which would make the counting argument easier to verify.","section":"Section 4.2, Eq. (17)"},{"comment":"The statement that the derivative of Delta_h with respect to h_N is constant is correct given Eq. (16b), but it would be clearer to state that this constancy relies on the measured heads y_h and source heads h_s being fixed inputs in the inversion.","section":"Section 5.1, Eq. (22b) and Eq. (23)"},{"comment":"The treatment of roughness values as absolute values is described only informally in Remark 4; the manuscript should specify exactly which components are reflected and how this reflection affects the Jacobian computation in later iterations.","section":"Algorithm 1, line 17 and Remark 4"},{"comment":"The manuscript contains numerous typographical and formatting errors, for example 'ndefine' in the Notation section and inconsistent units in the presentation of v(x) in Table 4; a careful proofread would improve readability.","section":"General presentation"}],"recommendation":"major_revision","confidential_remarks":"The core formulation is plausible and the paper may become publishable after (i) repairing or replacing the line-search/descent argument for the rectangular system, (ii) providing conditions or numerical verification for Assumption 5, and (iii) expanding the evaluation beyond a single noise-free example. The novelty relative to existing explicit calibration literature is asserted rather than demonstrated; the authors should position their contribution more carefully against prior work on explicit roughness calibration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know before reading.\n\nThe useful part is the explicit Colebrook-White turbulent flow expression (12) and the calibration problem (16) built on it. That formulation, together with the counting rule nm,min = ceil(n_l/n_p) in (17), is new and cleanly stated. The paper also does well in making its assumptions explicit, including the unrealistic but clearly acknowledged assumption that nodal consumption is known, and in restricting the scope to full turbulent flow. The one detailed example is reproducible in principle and shows the method can work when the rank condition holds.\n\nThe weak part is the solver analysis. The paper claims the algorithms reliably find the real root, but the supporting calculation in Eq. (26) is wrong for the overdetermined case. The Jacobian is rectangular, the Newton step uses the left inverse, and the directional derivative along that step is -sign(f)^T J (J^T J)^-1 J^T f, not -sign(f)^T f. The line search therefore rests on a false premise. The algorithm may still work, and the example suggests it does, but the convergence proof as written does not establish the claimed reliability. The stress-test note lands.\n\nOther soft spots are proportional. Assumption 5, the linear independence of measurement sets, is asserted rather than proved, and it is genuinely load-bearing: the counting rule is only necessary, not sufficient. The paper does not verify the rank condition for its example beyond the successful run. The evaluation is a single noise-free simulation generated with the authors' own dynamic model, whose equilibrium is equivalent to the steady-state equations being inverted; that is not circular numerically, but it does not test generalizability. There is no noise study, no comparison against existing implicit or explicit calibration methods, and no shipped code. Those are standard gaps, not fatal flaws in the problem formulation itself.\n\nBottom line: the mathematical setup and the counting argument are worth publishing, and the explicit turbulent flow formulation is a genuine contribution. But the reliability claim needs a corrected descent analysis, rank verification, and noise experiments before it can be taken seriously. The paper deserves a serious referee, not a desk reject. I would send it to review with the expectation of heavy revision.\n\nFor a reader: someone working on model calibration or leakage detection in water networks, or on identifiability of inverse problems with redundant measurements, will get value from the formulation and the counting rule. I would not cite the solver, but I would cite the explicit flow expression if I needed it.","headline":"Useful new formulation for per-pipe roughness identification from pressure measurements, but the convergence proof has a genuine flaw and the reliability claim outruns the evidence.","tokens_in":19742,"tokens_out":3101,"would_cite":true,"duration_ms":32925,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65H10","93B30"],"pacs":[],"model":"deepseek-v4-flash","headline":"Under steady fully turbulent flow, individual pipe roughnesses in a water distribution network can be recovered by inverting the hydraulic equations using pressure measurements at a subset of nodes, provided enough independent measurement…","keywords":["Roughness calibration","Water distribution networks","Parameter identification","Colebrook-White","Darcy-Weisbach","Hydraulic friction parameters","Newton-Raphson","Steady-state hydraulics"],"falsifier":"Compute the numerical rank of $J(x^*)$ from equation (20) at the true roughnesses and the measurement sets in Table 2; if the rank is below $n_\\ell+n_m(n_j-n_p)$, then some nonzero direction $d$ satisfies $Jd=0$ and two different roughness/pressure configurations produce identical measurements, disproving unique identifiability for that configuration. A simpler experimental check is to run Algorithm 2 with measurement sets that differ only by a uniform scaling of all consumptions: if the returned roughnesses vary with random starts, the sets are not sufficiently independent.","tokens_in":18668,"feed_emoji":"🚰","tokens_out":11652,"duration_ms":164200,"temperature":0.7,"pith_summary":"Working in the steady state and assuming the flow in every pipe is fully turbulent, the paper claims that individual roughness values for all pipes of a water distribution network can be solved for directly from the hydraulic network equations, using pressure measurements at a subset of nodes and known nodal consumption. The key step is to rewrite the turbulent head-loss relation as the explicit Colebrook–White flow function $f_t(\\epsilon,\\Delta h)$, stack the nodal mass-balance equations over several loading conditions, and count unknowns: at least $n_m=\\lceil n_\\ell/n_p\\rceil$ independent measurement sets are needed. The authors propose a modified Newton–Raphson scheme with an $\\ell^1$-norm backtracking line search and multi-start initialization, and show on an eight-pipe example that all eight roughnesses and the unmeasured pressure heads are recovered, with the largest roughness error about 6% in the noise-free case. If correct, this makes per-pipe roughness calibration from a sparse pressure-sensor layout feasible in the fully turbulent regime, which would improve model-based leakage detection and localization.","feed_headline":"Counting rule recovers per-pipe roughness from a few pressure sensors","feed_subtitle":"With enough independent steady-state pressure sets, a Newton solver returns every pipe's roughness and unmeasured heads.","key_machinery":"The load-bearing object is the stacked equation set (16). For each measurement set $i$, it combines nodal mass conservation $A x_Q(\\boldsymbol\\epsilon,h_N^{(i)})=\\bar q^{(i)}$ with the head-loss relation $\\Delta h^{(i)}=\\tilde C_s h_s^{(i)}-A^T C_h^T y_h^{(i)}-A^T\\bar C_h^T h_N^{(i)}-A^T z$ and defines each pipe flow by the explicit turbulent Colebrook–White function $f_t(\\epsilon_j,\\Delta h_j)$ from equation (12). The stacked residual $f(x)$ has a tall rectangular Jacobian $J(x)$ whose block structure consists of repeated incidence matrices $A$ multiplying blocks that are diagonal in the roughness derivatives and dense in the unmeasured-head derivatives. The paper's counting rule $n_{m,\\min}=\\lceil n_\\ell/n_p\\rceil$ is what balances equations against unknowns, and the modified Newton algorithm uses the left inverse $(J^T J)^{-1}J^T$ with a backtracking line search based on the $\\ell^1$ norm, plus physical box constraints on roughness and unmeasured heads to steer iterates back to the real root.","core_discovery":"On its own terms, the paper's central claim is that the inversion problem is solvable as stated: with sufficiently independent steady-state measurement sets, solving the stacked equation system (16) recovers the true roughness vector $\\boldsymbol\\epsilon\\in\\mathbb{R}^{n_\\ell}_{\\ge 0}$ together with all unmeasured pressure heads $h_N^{(i)}$ in every measurement set. The discovery is that the implicit Colebrook–White friction law can be converted into an explicit flow function $f_t$, which makes the Jacobian of the stacked residual computable and enables a Newton-type solver. The authors show that the minimum number of measurement sets is $n_{m,\\min}=\\lceil n_\\ell/n_p\\rceil$, independent of the number of nodes, and they provide two algorithms: a modified Newton method with step-length control, and an outer multi-start procedure that reinitializes only the roughnesses that leave their physical range. In the noise-free three-cycle simulation, the algorithms converge to the real root; the authors state in the conclusion that the algorithms find the real root reliably whenever independent measurement-sets are available.","pith_inferences":["The paper does not prove that Assumption 5 holds for a given network; before relying on the calibration, an operator should check the numerical rank of the stacked Jacobian for the chosen sensor placement and load patterns.","Because nodal consumption is assumed known, field application will likely be limited by consumption estimation error; fireflow tests, in which hydrant outflows are deliberately measured, are the natural setting where this assumption is approximately met.","The block structure of the Jacobian suggests each measurement set contributes an independent block coupled only through the shared roughness vector, so the solver could be parallelized or solved with a Schur-complement reduction on large networks.","A deterministic alternative to the random multi-start outer loop, such as continuing from one load scenario to the next by gradually changing consumption, might recover the global root with fewer function evaluations."],"forward_implications":["The counting rule $n_{m,\\min}=\\lceil n_\\ell/n_p\\rceil$ means the required number of loading scenarios depends only on the ratio of pipes to pressure sensors, not on the number of nodes.","With that many sufficiently independent measurement sets, all $n_\\ell$ roughnesses and all $n_m(n_j-n_p)$ unmeasured pressure heads are identifiable from pressure measurements alone.","Using the explicit turbulent flow function $f_t$ avoids iterating on the implicit friction factor inside each function evaluation, which keeps the Jacobian construction tractable.","Adding more measurement sets beyond the minimum, with clearly separated measured heads, improves robustness to noise and helps distinguish the true root from local minima, at the cost of a larger equation set.","Extending the method to real networks requires a sufficiently smooth explicit model for laminar and transitional flows, as the paper states in its outlook."],"supporting_citations":[{"why":"Supplies the steady-state network equations $Ax_Q=\\bar q$ and $A^T(h+z)=\\tilde C_s h_s-h_{\\rm loss}(x_Q)$ that the inversion problem is built on.","marker":"[8]"},{"why":"Provides the dynamic network model whose steady-state equilibrium is proven equivalent to the solution of (2), and is used to generate the independent measurement-sets.","marker":"[15]"},{"why":"The companion dynamic model for smart water networks used to produce the simulated consumption and pressure scenarios.","marker":"[16]"},{"why":"Defines the Reynolds-number regime boundaries and the fluid-mechanics background for the turbulent-flow assumption.","marker":"[21]"},{"why":"Provides the implicit friction-factor relation that is the basis for the explicit turbulent flow function $f_t$ used in the equation set.","marker":"[22]"},{"why":"Supplies the backtracking line-search and polynomial interpolation strategy used for step-length control in Algorithm 1.","marker":"[28]"}],"fun_headline_variants":["Few pressure gauges reveal each pipe's roughness","Sparse pressure sets unveil every pipe's roughness","Full turbulent: pipe roughness from a few pressure sets","Newton solver recovers all pipe roughness from sparse gauges"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on Assumption 5, that the available measurement sets are \"sufficiently linearly independent\" so the stacked Jacobian has full column rank; this is asserted rather than proved, and the counting rule alone does not guarantee it for a given topology, sensor placement, or load pattern.","fun_headline_variants_meta":{"raw":{"variants":["Few pressure gauges reveal each pipe's roughness","Sparse pressure sets unveil every pipe's roughness","Full turbulent: pipe roughness from a few pressure sets","Newton solver recovers all pipe roughness from sparse gauges"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000892,"raw_usage":{"total_tokens":3815,"prompt_tokens":879,"completion_tokens":2936,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":495,"completion_tokens_details":{"reasoning_tokens":2873}},"tokens_in":495,"tokens_out":2936,"duration_ms":538157,"temperature":1.0,"reasoning_tokens":2873,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:56:02.578445+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the numerical rank of $J(x^*)$ from equation (20) at the true roughnesses and the measurement sets in Table 2; if the rank is below $n_\\ell+n_m(n_j-n_p)$, then some nonzero direction $d$ satisfies $Jd=0$ and two different roughness/pressure configurations produce identical measurements, disproving unique identifiability for that configuration. A simpler experimental check is to run Algorithm 2 with measurement sets that differ only by a uniform scaling of all consumptions: if the returned roughnesses vary with random starts, the sets are not sufficiently independent.","supporting_citations":[{"cited_title":"Todini, S","cited_arxiv_id":null,"evidence_quote":"Supplies the steady-state network equations $Ax_Q=\\bar q$ and $A^T(h+z)=\\tilde C_s h_s-h_{\\rm loss}(x_Q)$ that the inversion problem is built on."},{"cited_title":"Kaltenbacher, M","cited_arxiv_id":null,"evidence_quote":"Provides the dynamic network model whose steady-state equilibrium is proven equivalent to the solution of (2), and is used to generate the independent measurement-sets."},{"cited_title":"Kaltenbacher, D","cited_arxiv_id":null,"evidence_quote":"The companion dynamic model for smart water networks used to produce the simulated consumption and pressure scenarios."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Reynolds-number regime boundaries and the fluid-mechanics background for the turbulent-flow assumption."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the implicit friction-factor relation that is the basis for the explicit turbulent flow function $f_t$ used in the equation set."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the backtracking line-search and polynomial interpolation strategy used for step-length control in Algorithm 1."}],"review_version":1}