acceptodds
Under review as a conference paper at ICLR 2027

What Does a Thermal Surrogate Lose Without the Flow Field? Forecasting Valve-Induced Transients in Liquid-Cooled Datacenters with Graph Neural ODEs

13 views

Abstract

Learned surrogates of thermal-fluid plants are usually trained and evaluated with the flow field given. In a liquid-cooled datacenter, closing one cooling channel's valve pushes its flow onto the sibling channels within seconds, and a deployed surrogate rarely observes that post-event flow. We measure what this costs and whether the missing flow can be supplied. On a transient simulator of the Frontier supercomputer's cooling plant, we train graph surrogates under four information regimes: true flow, flow frozen at the forecast start, a closed-form branch-conductance estimate computed from pre-event measurements and the valve command, and frozen flow plus the command given to the network. We evaluate each on 49 held-out closures. Without post-event flow, the sibling-channel error of a flow-structured graph neural ODE over the first minute rises about 32-fold (from 0.032 to 1.013 K), two baselines fail in the same way, and every model predicts the wrong direction of change on the closed channel in at least 47 of the 49 events. Event-locked controls show that the gap appears only at the event. The closed-form estimate, a standard virtual flow meter, recovers 99% of the gap. Given the command instead, the network recovers 85% on the closed channel but only 6% on its siblings: it learns what a valve does to its own channel, not how the flow redistributes to its neighbours.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
graph neural ODEs, learned surrogates, thermal-fluid systems, liquid-cooled datacenters, partial observability, virtual flow metering, physics-informed machine learning, message passing, spatio-temporal forecasting, digital twins

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_DiFCOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This is an empirical study where the authors adopt an established graph neural ODE approach, namely AdvGNODE, to predict temperature changes in server liquid cooling. They study how missing flow information affects the predictions and whether estimating this flow can improve the results.

Critical strengths

None

Critical weaknesses

As this is an empirical paper, in the assessment I will evaluate the paper based on the importance of the problem, the practical impact of the results, and whether the approach can be useful for related problems. I believe the paper does not sufficiently address these points:

  • Regarding the problem, although temperature prediction for liquid cooling is interesting, the authors only evaluated their approach in simulation; real-world validation or deployment would strengthen the paper significantly. Also, although the paper is motivated by control, in the end, the authors only evaluate forecasting. This makes it less convincing that the results would be useful for control problem in practice.

  • Regarding the approach, this seems to be a straightforward application of an existing method to their problem. The results show how missing flow affects prediction in their simulation, but these insights seem limited to their specific problem, and what values could they bring to broader related problems remains unclear.

Questions for authors

None

1 comment
AuthorsOpenReview ↗

Response: simulation scope, forecasting vs control, novelty.

Thank you for the review.

Simulation only. A real limitation, stated in Section 7. Branch-level flow measurements around valve events are not publicly available for the real Frontier system, so we use the public transient model of its cooling plant (ExaDigiT, distributed with LC-Opt), which makes the study reproducible. We now also report a second plant with nonlinear (equal-percentage) valves, where the same conclusions hold: ESTIMATED is within 0.001 K of TRUE-FLOW on the siblings at 60 s, and FROZEN and COMMAND are about 20× worse.

Control versus forecasting. The use case is forecasting inside a controller, before an action is taken. We now show that the forecast error changes decisions: ranking the 49 test closures by the predicted temperature change, Spearman correlation with the truth on the struck channel is 0.998 (ESTIMATED), 0.931 (COMMAND) and −0.16 (FROZEN); on the siblings 0.983, 0.435 and 0.611.

Novelty and broader value. Please see the general response. We do not claim a new architecture; the new experiments show what a network needs to learn the redistribution, which applies to any flow network with unmeasured actuation.

Reviewer_MLgDOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies graph-based surrogates for thermal fluid plants, specifically after a valve closure event that induces flow changes. The core contribution is its study of the information needed by off-the-shelf surrogates (at training and inference time) to produce reliable predictions after the valve closure. Experiments train and evaluate multiple surrogate modeling approaches and utilize four flow information conditions, ultimately finding that surrogate success depends on either ground truth flow fields or estimates thereof (from a closed-form model).

Critical strengths

The paper is well written and addresses an interesting failure mode of off-the-shelf surrogate models.

Experiments are carefully designed and effectively test the submission’s core hypotheses, providing insight into the information that thermal surrogates require to maintain accuracy after a valve closure.

Critical weaknesses

My main concern relates to significance. The submission is narrowly focused on the response of existing surrogate models to a specific event (a valve closure) within a specific application (thermal fluid modeling). As written, I think readers will wonder whether the conclusions drawn from this study could be informative beyond this setting.

Questions for authors

Towards broadening the significance, are there similar applications that would benefit from the insights developed here? E.g., how do the results here translate to the hydronic and other systems discussed in the related work section?

Alternatively, could the flaws of the surrogate models be addressed somehow, perhaps through the approaches suggested on line 432, such that the redistribution is learned? Addressing this would augment the fundamental machine learning contribution of the submission and thus its broader significance.

3 comments
AuthorsOpenReview ↗

Response: learnability of the redistribution, significance beyond this setting.

Thank you for the review and for two concrete suggestions; we took up both.

Can the redistribution be learned (line 432)? Yes, under conditions we can now state. Routing the command to the whole cabinet recovers about 90% of the sibling gap (vs 6% for the local command); a learned flow predictor trained on the 392 training closures matches the closed-form estimate within 0.005 K. Routes without that explicit input do less: bidirectional messages with a local command recover almost nothing more (sibling 0.957 vs 0.972 K), and STGCN with the command about half (0.515 K). The general lesson: hydraulic redistribution is set by network-wide constraints at the moment of actuation, while thermal transport is local; a message-passing surrogate built on advective locality learns the second but not the first from a local command. Supplying the global part, by a closed-form solve or by routing the command, fixes it, and the closed form does so most accurately, including on actions absent from training. This turns the paper from a diagnosis into a diagnosis plus what fixes it.

Significance beyond this setting. The setting is any learned surrogate of a network in which an actuator redistributes a conserved flow that is not measured after the action: district heating and cooling, hydronic HVAC, water distribution. Section 7 now states this setting. New evidence of breadth: a second plant with nonlinear valves (ESTIMATED recovers 100% of the sibling gap there, COMMAND none) and actions never seen in training (ESTIMATED within 0.05 K of TRUE-FLOW on six new action types, while command models degrade; Table R5).

Presentation. The Introduction now states the use case and what each regime measures; results lead with absolute errors, and the 32× ratio appears once in the main text; "99%" is scoped to the 60 s after the event throughout; the new results sit in Sections 6.2–6.3 (one table, one figure) with details in Appendix D.

We hope these address your concerns. If they do, we would be grateful if you would consider updating your assessment, and we are happy to answer further questions.

I thank the authors for addressing the questions in my review but am still worried about whether this contribution will be seen as significant at ICLR. Accordingly, I'll maintain my rating while noting to the AC that I think this is a well done study -- for me, it is (however) too focused on addressing a failure mode within an application of ML, as opposed to advancing ML more broadly (and thus whether it belongs in ICLR, as opposed to a different venue, is unclear to me).

Thank you for the careful review and for your kind note to the AC. We understand the concern about fit, and we respect it. We would only add, for the AC's consideration, that we submitted under ICLR's area on applications to the physical sciences, since we see the finding as one about learned surrogates more generally: they can fail when an actuation changes an input they are not given, beyond this particular plant. We appreciate your engagement throughout.

Reviewer_qMcDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

THis paper presntes learned temperature surrogates of the Frontier liquid-cooling plant (ExaDigiT/LC-Opt FMU) through single-channel valve closures. It trains graph surrogates under four information regimes: TRUE-FLOW, FROZEN, ESTIMATED (a closed-form branch-conductance estimate), and COMMAND (valve opening as an extra input). The results: Withholding post-event flow raises AdvGNODE’s sibling-channel RMSE about 32-fold (0.032 -->1.013 K). 2) Every model gets the sign wrong on the struck channel in at least 47 of 49 closures. The closed-form estimate recovers 99% of the gap. COMMAND recovers 85% on the struck channel but only 6% on the siblings.

Critical strengths
  • The within-rollout event-locked control (Fig. 2, Table 8), the same-weights swap (Table 9), and the setpoint contrast class are good causal-attribution design.
  • Training and evaluating in the same regime avoids a train/test mismatch confound.
  • The graph is derived from port connectivity and verified by mass balance.
  • Reporting wrong-sign counts and ratios (Table 2) is more informative than RMSE alone. THis paper is open about the post-hoc diagnostics and the limited seeds.
Critical weaknesses
  • A network told nothing about an event cannot forecast its effect. With the event at t0 and frozen flow, FROZEN has no signal. Its output (a mild cooling everywhere, about −0.46 K) is just the conditional mean.
  • 32x figure is a ratio over a very small denominator (note: TRUE-FLOW is an oracle for all future flow). It is measured at windows starting exactly at the event, the worst case, and says little about deployment cost.
  • it seems FROZEN is a strawman, Actuation commands are known to whoever moves the valve, and action-conditioned dynamics models are standard (model-based RL, MPC for HVAC, neural CDEs). The premise that "the flow is not observed" is only interesting if the command is also unavailable, and COMMAND shows it is available.
  • Sibling advective weights w_e use frozen flows, and the command enters only the struck node. The architecture have almost no route to representing redistribution.
  • The estimator is fit to the same simulator whose hydraulics are linear (branch R² > 0.9999, valve coefficient equal to nominal to machine precision). Its median sibling-flow-change error is 1.3%, hence, it effectively feeds back the true flow.
  • In Table 8, it hints at degradation. At δ = 110, ESTIMATED−TRUE is 0.107 K against 0.232 K for FROZEN−TRUE over the first 10 post-event seconds, only about 54% recovery, though this is confounded by pre-event drift. The “99%” is specific to event-anchored 60 s windows.
Questions for authors
  • clarifiy the novelty. AdvGNODE is not a new architecture as methioned in Sec. 4.
  • What happen if all three openings go to every cabinet node, or the model has global/bidirectional message passing? Does COMMAND then close the sibling gap?
  • How does THe learned flow predictor trained on the 392 closures compare with the closed-form estimator?
  • How much noise or bias in Δṁ can ESTIMATED tolerate before the gain disappears? Add an equal-percentage valve variant.
  • How events placed within training windows? Would FROZEN predict heating if event windows were distributed differently?
  • What is the deployment-style cost: mixed-offset windows, a realistic event rate, horizons to 300+ s?
14 comments
AuthorsOpenReview ↗

Response (1/2): deployment cost, command routing, nonlinear plant

Thank you for a detailed review. You read the appendix closely, and several of your points changed the paper. Point by point:

1. FROZEN has no event signal, so its output is a conditional mean. Agreed, and that is the mechanism we document. FROZEN is the floor, not the finding. What the paper measures is how large the cost is, where it lands, and that a near-mean forecast predicts the wrong direction on the struck channel in at least 47 of 49 closures, which aggregate metrics hide. The Introduction now says this explicitly.

2. The 32× ratio and the worst-case windows. Agreed on both. We now lead with absolute errors and report the ratio once in the main text. You asked for the deployment cost: over 2,000 windows at all offsets and the natural event rate, sibling RMSE is 0.017 / 0.033 / 0.207 / 0.205 K for TRUE-FLOW / ESTIMATED / FROZEN / COMMAND. Averaged over all nodes it is much smaller (0.066 / 0.072 / 0.088 / 0.071 K), so you are right that the plant-wide average cost is modest; it is concentrated on the affected channels in the first minute after an event, which is when control decisions are made (see the closure ranking in point 3–4 below).

3–4. FROZEN as a strawman; the command has no route to the siblings. This was the right objection, and we tested it. With all three cabinet openings at every channel node (your Q2), sibling RMSE at 60 s drops from 0.972 to 0.135 K (3 seeds): the paper's COMMAND result was largely a routing limitation, and we now say so plainly.

Three things still hold. (i) With the route in place, the command model's sibling error is still about 3× ESTIMATED's (0.042 K). (ii) The routing it needs is the hydraulic topology the estimator already uses, and the estimator is a small physical model rather than a learned mapping: one parameter per block is fitted on the training closures, and without it (the cabinet's total flow assumed unchanged) sibling RMSE is still 0.108 K (Table R3). (iii) The estimate holds on actions absent from training, where the command models degrade (point 5).

In deployment-style windows the paper's COMMAND model is no better than FROZEN on siblings (point 2), and it misranks sibling effects across closures (Spearman 0.44 vs 0.98 for ESTIMATED). The cabinet-command model gets directions right and ranks sibling effects nearly as well (Spearman 0.946 vs 0.983), but its error is larger and grows faster with the horizon: struck RMSE 0.235 vs 0.088 K for ESTIMATED over 60 s, and 0.91 vs 0.11 K over 120 s (Appendix D).

Your other suggestion, bidirectional message passing, does not close the gap: with messages in both directions and the local command, sibling RMSE at 60 s is 0.957 K (vs 0.972 K) and the predicted sibling change is still about a third of the true one (predicted / true ratio 0.37, vs 0.35 for the paper's COMMAND in Table 2). STGCN with the command reaches 0.515 K (vs 1.043 K without it). All three variants remain worse than ESTIMATED on both the struck and the sibling channels (Table R1; for the cabinet command +0.081 K and +0.082 K, Holm p 8.4e-3 and 1.4e-12). We now cite action-conditioned dynamics models (model-based RL, MPC for buildings, neural CDEs) in Section 2 and position our COMMAND regimes among them.

5. The estimator is fit to a linear simulator. Agreed for the submitted results. We now measure the tolerance (general response, point 2): RextR_{\text{ext}} halved or doubled, lognormal noise on it, and realistic sensor noise all stay far below FROZEN; a wrong valve law fails, and is detectable from data. On a second plant with equal-percentage valves, where the estimator is no longer fit to linear hydraulics, ESTIMATED still matches TRUE-FLOW: sibling RMSE at 60 s 0.068 vs 0.069 K (not distinguishable after Holm; 3 seeds), against 1.37 K for both FROZEN and COMMAND; the estimator's median sibling flow-change error is 0.9%. So the gain does not depend on the plant being linear; it depends on knowing the valve law, which is identifiable from logged data. On six action types never seen in training (re-openings, partial closures, two channels or two blocks at once, low sibling openings, close-then-reopen; 50 new rollouts each), the estimator keeps its accuracy (sibling median error 0.5–2.0%) and ESTIMATED stays within 0.05 K of TRUE-FLOW on the siblings, while the command-conditioned models degrade wherever the action differs from training closures: sibling RMSE at 60 s with low sibling openings is 0.175 K (ESTIMATED), 1.09 K (cabinet command) and 3.11 K (COMMAND); for two channels closing together 0.170, 0.527 and 2.50 K. On re-openings the cabinet-command model gets the sibling direction right in 31% of events vs 99% for ESTIMATED. This is the practical difference between learning the command's effect from examples and computing it from structure. Re-opening a valve from nearly closed (opening below 0.1) is hard for every model, TRUE-FLOW included; we report it separately.

AuthorsOpenReview ↗

Response (2/2): answers to Q1–Q6, and continued part 6 from response (1/2)

6. Table 8, δ=110\delta = 110: about 54% recovery in the first 10 s. Your reading is correct, and the paper now says so (Appendix D, "The first seconds after the event"). Net of the pre-event gaps (0.031 K ESTIMATED, 0.039 K FROZEN), first-10 s recovery on the siblings is 61% (seed range 45–76%) in Table 8's metric (mean absolute error), 67% in RMSE; on the struck channel it is 89%. The estimate is the quasi-steady flow 20–30 s after the closure, applied as a step, so its error is largest in the first seconds. Figure 2 now also shows the ESTIMATED gap per second for every δ\delta (new panel c); it stays at or below 0.13 K. "99%" now refers only to event-anchored 60 s windows.

Q1. Novelty. We do not claim a new architecture. The contributions are (i) a measurement of what an unobserved flow redistribution costs a learned surrogate, and where; (ii) the comparison of a closed-form estimate against command conditioning, now including cabinet-level routing, bidirectional messages, STGCN and a learned flow predictor; (iii) the event-anchored protocol. The new results add a clear statement of what it takes: the redistribution must be supplied, either as an estimate or as a routed command, and the estimate does it more accurately, including on actions absent from training.

Q2. Please see points 3–4.

Q3. Learned flow predictor. A small MLP trained on the 392 training closures with the estimator's exact information budget, fed to the ESTIMATED checkpoints without retraining: forecast RMSE, averaged over predictor seeds, within about 0.005 K of the closed form on every row (sibling 60 s 0.041 K), with lower flow error (sibling median 0.4–0.5% across predictor seeds vs 1.23%). With 25% of the closures, struck 60 s error is about three to four times larger (0.28–0.37 vs 0.08–0.09 K; Table R6). Redistribution can be learned when the flow is the target, but it needs the training events.

Q4. Noise or bias in Δm˙\Delta\dot{m}; equal-percentage valves. Please see point 5, Figure R1 and Table R3.

Q5. Event placement in training windows. Each epoch draws 12,000 windows, a quarter with an event (Section 5), placed uniformly within the 32 s window: over all 300 training epochs (3 seeds × 100), each quarter of the window holds 24.9–25.1% of the event windows (mean position 0.52 of the window; 82% closures, 18% setpoint changes). Appendix D gives the numbers and Figure R2 shows the shares. Under FROZEN nothing in the input says that an event happened or which channel was struck, so no placement can make the model predict struck-channel heating; it can only move its conditional mean.

Q6. Deployment-style cost and horizons to 300+ s. Mixed offsets: see point 2. Horizons: sibling RMSE ESTIMATED 0.141 / 0.426 / 1.195 K at 120 / 240 / 480 s vs FROZEN 1.168 / 1.268 / 1.747 and COMMAND 1.190 / 1.370 / 1.660; struck stays strong (480 s: 0.97 vs FROZEN 9.67, COMMAND 5.25 K). The sibling advantage fades, and we checked why. No command or input changes after a closure; the logged flows drift the same way in re-runs without the event; TRUE-FLOW sees that future drift (your "oracle for all future flow"). An ideal one-shot estimate, the logged flow 20–30 s after the event, scores the same as ours within 0.003 K at every horizon. So the fade is a limit of any estimate made at the event, not of our estimator, and we state it as a limitation.

We hope these address your concerns. If they do, we would be grateful if you would consider updating your assessment, and we are happy to answer further questions.

THe authors have substantially provided verbal explanations. They report additional experiments addressing the specific questions.however, the novelty is still somewhat limited The authors explicitly acknowledge that they do not propose a new architecture. the contributions in this paper are now framed around: Measuring the cost of unobserved flow redistribution, and Comparing physical estimation against command-conditioned learning.

This is a reasonable repositioning, but it raises a fundamental question as follows: What is the actual scientific contribution beyond demonstrating that a physically informed estimator outperforms an inadequately informed neural surrogate?

Reviewer_SFzQOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates the impact of unobserved post-actuation flow redistribution on temperature forecasting in liquid-cooled datacenters. Using a simulator of the Frontier cooling plant, it evaluates graph-based thermal surrogates under four input regimes: (i) true flow, (ii) frozen pre-event flow, (iii) hydraulically estimated flow, and (iv) frozen flow supplemented with valve commands.

The main contributions are: (i) an empirical characterization of the resulting forecast degradation across three surrogate architectures, (ii) an event-aligned evaluation protocol that reveals localized failures obscured by aggregate metrics, and (iii) a controlled comparison of physics-based flow estimation with direct conditioning on valve commands.

Experiments on 49 held-out valve closures demonstrate substantial forecasting errors when post-event flow changes are unobserved. This includes incorrect predictions of the direction of temperature change in the actuated channel.

Critical strengths

[S1] The study provides convincing experimental results demonstrating that missing post-actuation flow information drives forecasting failures.

The experimental design helps isolate the effect of missing flow information. Firstly, each model is trained and evaluated within the same information regime, reducing confounding from training-deployment mismatch. Next, event-aligned comparisons show that errors increase sharply after valve closure, while changing flow inputs without changing model weights reproduces the failure. Finally, together with consistent results across three architectures, these controls strengthen the conclusion that missing flow information drives the observed degradation.

[S2] The evaluation exposes physically meaningful failures that existing aggregate forecasting metrics can obscure.

Plant-wide error metrics can obscure physically meaningful failures by averaging errors in affected channels with accurate predictions elsewhere. By evaluating components separately and checking the direction of temperature change, the paper shows that acceptable overall scores can conceal predicted cooling in a channel that actually heats, alongside underestimated cooling in neighboring channels. The comparison also helpfully shows the importance of representing flow redistribution: (i) local valve commands largely recover the actuated channel’s response, whereas (ii) hydraulic estimates additionally recover the response of neighboring channels.

Critical weaknesses

I lean toward acceptance because the controlled experiments establish a useful and interpretable failure mode for thermal surrogates. Nonetheless, the restricted command baseline and idealized hydraulics limit the breadth of the conclusions.

[W1] The provided baseline (the COMMAND variant of AdvGNODE) is restricted in its ability to capture flow redistribution between sibling channels, making it unclear whether the advantage of hydraulic estimation reflects the value of explicit physics or the baseline’s architectural limitations. In other words, it leaves the benefit of explicit hydraulic estimation over a more suitable command-conditioned architecture unresolved.

Each valve command is supplied only to its own channel, and messages travel in the direction of coolant flow. This prevents the command from directly informing sibling channels about the resulting flow redistribution. The low recovery may therefore reflect how the baseline routes information. Although the authors acknowledge this limitation, giving all channels access to the cabinet’s valve commands would provide a stronger comparison and clarify the added benefit of explicit hydraulic estimation.

[W2] The near-complete recovery is demonstrated only under favorable hydraulic modeling assumptions, leaving its robustness to model mismatch and measurement noise uncertain.

The experiments use a single simulator with linear valves, noiseless measurements, and a hydraulic estimator that closely matches the underlying branch physics. The reported 99% recovery therefore establishes effectiveness under this particular scenario and accurate flow estimation, while real-world robustness to calibration error, nonlinear valve behavior, or sensor noise remains untested. A sensitivity analysis linking flow-estimation errors to temperature-forecast errors would help to strengthen the practical significance of the findings and make the contribution stronger and more applicable.

Questions for authors

[Q1] Would broadcasting valve commands to all channels within a cabinet improve sibling-channel forecasts, and does hydraulic estimation retain its advantage over this stronger baseline?

[Q2] How sensitive is the reported 99% recovery to errors in estimated flow, such as those arising from sensor noise, imperfect calibration, or nonlinear valve behavior?

[Q3] Could the authors clarify whether the base model changes both advective and thermal-message weighting, and provide an ablation that separates their contributions to forecast accuracy?

1 comment
AuthorsOpenReview ↗

Response: command broadcast to all channels, sensitivity to estimation error, ablation.

Thank you for the positive and precise review.

W1 / Q1. Command broadcast to all channels. We ran it. With every channel node receiving all three of its cabinet's commanded openings, sibling RMSE at 60 s is 0.135 K (3 seeds) vs 0.972 K when the command enters only the struck channel, and 0.042 K for ESTIMATED; struck 0.235 vs 0.846 and 0.088 K. Both differences from ESTIMATED are significant after Holm correction (sibling +0.082 K, p 1.4e-12; struck +0.081 K, p 8.4e-3). Its error also grows faster with the horizon: struck RMSE 0.91 K over 120 s vs 0.11 K for ESTIMATED (Appendix D). So your reading was right: most of the sibling gap was how the baseline routed information. Hydraulic estimation keeps about a 3× advantage on the siblings, and it holds on actions absent from training, where the cabinet command degrades (with low sibling openings 0.175 vs 1.09 K; Table R5). We also ran the two alternatives (3 seeds each). Messages in both directions with the local command: sibling 0.957 K, struck 0.800 K, against 0.972 / 0.846 K for the local command alone (0.015 K lower on the siblings; the struck difference is not significant). STGCN with the command: sibling 0.515 K vs 1.043 K for STGCN without it, struck 0.544 vs 3.99 K; better, but about 12× ESTIMATED's sibling error. Table R1 (main text); per-variant results on unseen actions in Table R5.

W2 / Q2. Sensitivity to estimation error, noise and nonlinear valves. We added the analysis you suggest (Figure R1 and Table R3); see the general response, point 2, for the numbers. In short: recovery stays above 93% when RextR_{\text{ext}} (the fitted resistance that stands for the rest of the plant, the estimator's least certain parameter) is halved or doubled, and at about 84% or more for meter, flow and temperature noise up to 5% / 5% / 0.2 K (sibling RMSE at most 0.19 K vs 1.01 K for FROZEN); only a wrong valve law breaks it, and that is detectable from data. On a second plant with equal-percentage valves, retrained in all four regimes, ESTIMATED recovers 100% of the sibling gap and 99.8% on the struck channel at 60 s (sibling 0.068 K vs 0.069 K TRUE-FLOW and 1.37 K FROZEN, 3 seeds), while COMMAND recovers 61% on the struck channel and nothing on the siblings.

Q3. What base changes, and an ablation. Yes, base differs from AdvGNODE in two ways (Section 4): flow enters the advective MLP as an input instead of scaling its output, and thermal messages are unweighted. We trained the two intermediate variants under TRUE-FLOW and FROZEN, 3 seeds each. The thermal-message weighting carries the difference. With flow entering as an input (as in base) but thermal messages weighted (as in AdvGNODE), the model matches AdvGNODE under TRUE-FLOW (sibling RMSE at 60 s 0.035 vs 0.032 K; none of the 8 Holm-corrected paired t-tests is significant). With flow-scaled advective messages but unweighted thermal messages, it performs like base (0.138 vs 0.134 K) and is worse than AdvGNODE in all 8 comparisons. Under FROZEN all four models are equal on the struck and sibling channels. This fits what the weighting is: a learned conductance on each thermal message that grows with the local flow (UA = softplus(θ)·w̄^α, α ∈ [0.4, 1]). So AdvGNODE's advantage still comes through the flow, via flow-dependent heat exchange rather than flow-scaled advection, and it disappears when the flow is frozen. We now say this in Section 6.3. The paper's conclusions are about information regimes and hold across all three architectures, so they do not depend on this. Section 4.3 states both differences; Appendix B said they differ only in the advective messages, and we corrected it. The ablation is in Section 6.3 and Table R2.

We hope these address your concerns. If they do, we would be grateful if you would consider whether they strengthen your assessment, and we are glad to clarify anything further.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.