{"id":"91c39f99-1fcf-4c1f-811a-7e315d1dd849","arxiv_id":"2607.28962","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A configurable four-state RC thermal model with deadband control simulates AI data center cooling load profiles, reducing MAE from 95.8 to 20.9 kW on Marconi100 telemetry.","lead":"This paper builds a state-space thermal model that turns IT workload and outside temperature into simulated cooling electricity use for hybrid air- and liquid-cooled data centers, and tests it against telemetry from the Marconi100 supercomputer. The model cuts mean absolute error from 95.8 kW to 20.9 kW versus a constant-COP baseline, but the parameters were tuned on the same data used for evaluation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Validation is in-sample only: parameters manually calibrated on the M100 data are evaluated against the same M100 reference, and Qliq is never defined, so the reported error reductions are not yet evidence of generalization.","rationale":"The reader identified the in-sample validation as the weakest assumption, and my independent read agrees: the paper's strongest empirical evidence is Table I, but both the proposed model and the baseline are calibrated and evaluated on the same M100 reference. This makes the reported error reduction consistent with overfitting, especially given the large number of tunable parameters. The reader also noted the missing definition of Qliq in the rationale, and I confirmed that Algorithm 1 step 11 references Qliq but no formula appears in the text or equations. This missing specification is an independent completeness issue, but it does not change the overall verdict: the paper is worth conditional acceptance provided the authors supply a held-out validation and a complete model specification. I considered whether the missing Qliq definition should push the verdict to REJECT or UNVERDICTED, but the released code may contain the missing formula, and the central modeling idea remains plausible. The concrete split-sample test would settle whether the in-sample metrics generalize; if they do not, the claim of a validated configurable model should be substantially revised. I therefore keep the reader's CONDITIONAL verdict unchanged.","tokens_in":4576,"tokens_out":5517,"duration_ms":52101,"concrete_test":"Using the released repository and the M100 telemetry, split the approximately 520 days into a calibration set (e.g., first 260 days) and a hold-out set (remaining 260 days, or use k-fold cross-validation by week). Re-calibrate all RC parameters, setpoints, deadband widths, and initial states only on the calibration set, then evaluate MAE/RMSE and daily statistics on the hold-out set. If the hold-out errors are materially worse than the in-sample errors (e.g., MAE above roughly 40 kW or close to the baseline error), the reported validation is overfit and the central claim needs to be weakened. Separately, confirm that the released code contains an explicit formula for Qliq(k); if not, the model specification is incomplete.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the proposed model reproduces dynamic cooling-electricity profiles far better than a constant-COP baseline. The load-bearing weakness is that the validation does not separate calibration from evaluation. Section II.B states that the model parameters are 'manually calibrated and iteratively adjusted' on the M100 data, and Section III.A states that the baseline COP is also selected by minimizing RMSE on the same measured reference. The results in Table I and Fig. 3 are then reported on that same reference. No held-out period, cross-validation, or second-facility test is presented. With a four-state RC network (four capacitances, five resistances), an additional rejection-state capacitance C5, plus setpoints, deadband widths, initial states, and rated capacities, the model has enough free parameters to fit one facility/time period. The reported MAE/RMSE reductions (95.80 -> 20.88 kW, 109.79 -> 27.27 kW) may therefore reflect in-sample calibration rather than a generalizable 'configurable' model. The reader's weakest assumption is correct on this point. Separately, the manuscript is also incomplete: Algorithm 1 step 11 says 'Compute Qliq(k)' but Qliq is never defined in the text or equations. Equation (7) depends on Qliq, so the model cannot be reproduced or adapted to another facility without an unpublished formula. This supports, rather than replaces, the need for a held-out evaluation and a complete specification.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a state-space thermal-dynamic model of a hybrid air/liquid-cooled AI data center, combining a four-state RC network for rack, air, mass, and coolant temperatures, a one-state heat-rejection node, and deadband-based on/off control to generate cooling electricity load profiles. The model is intended as a configurable, physically interpretable tool for long-duration power system studies. Validation against Marconi100 telemetry reports a reduction in MAE from 95.80 to 20.88 kW and in RMSE from 109.79 to 27.27 kW compared with a constant-COP baseline, together with improved reproduction of daily peak and variability statistics. Source code and sample data are released publicly.","tokens_in":5018,"tokens_out":5191,"duration_ms":47703,"significance":"If the reported performance were demonstrated out of sample, the paper would contribute a useful, computationally tractable alternative to constant-COP approximations for AI data center cooling load simulation. The public release of source code and sample data is commendable and supports reproducibility. However, the current evidence does not establish generalizability: the model parameters are manually tuned on the same facility/time series used for evaluation, the baseline is also optimized on the same reference, and key parts of the model specification (Qliq, Q_rej) are missing. These issues must be addressed before the central 'validated' claim can be accepted.","major_comments":[{"comment":"The validation is in-sample. Parameters are 'manually calibrated and iteratively adjusted' on the M100 data, and the constant-COP baseline is selected by minimizing RMSE on the same measured reference; Table I and Fig. 3 then report errors on that same reference. With dozens of free parameters (C1–C5, R1–R6, setpoints, deadband widths, rated capacities, initial states), the reported reductions (MAE 95.80→20.88 kW, RMSE 109.79→27.27 kW) may reflect fitting rather than predictive skill. A held-out period, cross-validation over days, or a second facility is needed to support the claim of validation, or the results must be explicitly reframed as in-sample reproduction.","section":"§II.B, §III.A, Table I"},{"comment":"The model is not fully specified. Algorithm 1 step 11 says 'Compute Qliq(k)', but no formula for Qliq is given anywhere in the text or equations. Equation (7) depends on Qliq, so without this definition the simulation cannot be reproduced or adapted to another facility. Similarly, Q_rej appears in Eq. (7) and is said to be introduced later, but Algorithm 1 only computes P_liq from the binary deadband state; the coupling between Q_rej, the deadband control, and P_liq is absent. Please add the missing equations and make the control-to-thermal coupling explicit.","section":"Algorithm 1, Eq. (7)"},{"comment":"The paper calls the model 'configurable' for other facilities, but no parameter-estimation or calibration procedure is provided. The only calibration described is manual, iterative adjustment within physically reasonable ranges, and parameter estimation is left to future work in the Conclusion. No identifiability or sensitivity analysis is given. A user cannot adapt the model to a new facility without significant ad hoc tuning. At minimum, the authors should provide a concrete calibration recipe (e.g., a least-squares or grid-search procedure) and assess how sensitive the simulated load profiles are to parameter values.","section":"§II.B, §IV"}],"minor_comments":[{"comment":"Typo: 'presents the a configurable' should be 'presents a configurable'.","section":"I"},{"comment":"The notation in Eqs. (2)–(3) has missing spaces (e.g., 'T rack,T air'), and matrices in Eq. (4) are typeset without visible bracket separators. Please fix the formatting for readability.","section":"II.A"},{"comment":"Figure citations are inconsistent: 'Figure 3 (a)' vs. 'Fig. 3(a)'. Use one style.","section":"III.B"},{"comment":"The baseline correlation Corr(Q_IT, P_cool) = 1.0 follows by construction from P_cool = Q_IT / COP; consider noting this explicitly to avoid implying an empirical correlation.","section":"Table I"},{"comment":"The paper says 'approximately 520 daily profiles' but does not state how many of these days, if any, were used during the manual calibration. Please clarify whether the statistical comparison is on the same period used for tuning.","section":"III.C"}],"recommendation":"major_revision","confidential_remarks":"The paper has a useful goal and provides code and data, but the central validation claim is not yet supported because the evaluation is in-sample and the model specification is incomplete. The missing Qliq/Q_rej equations are a straightforward fix, but the in-sample validation requires a genuine out-of-sample or cross-validated evaluation, or a substantial weakening of the claims. I recommend major revision, not rejection, because the underlying modeling approach is plausible and the authors can likely address the concerns with additional analysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is sensible: build a configurable RC thermal network with deadband-controlled cooling equipment, and use it to generate dynamic cooling-electricity profiles for power-system studies. That is a real gap, because time-synchronized cooling measurements are rare and constant-COP models miss the thermal inertia that matters for grid flexibility. The authors add a heat-rejection node to the standard four-state RC framework, which is a legitimate extension, and they release code and sample data, which is good practice.\n\nThe evaluation on Marconi100 telemetry shows the model tracking daily peaks and variability much better than a constant-COP baseline. But the validation is in-sample. Section II.B says the RC parameters, setpoints, deadbands, and initial states were \"manually calibrated and iteratively adjusted\" on the M100 data, and Section III.A says the baseline COP was also selected by minimizing RMSE on the same reference. No held-out period, no cross-validation, no second facility. With four capacitances, five resistances, both deadbands, both setpoints, rated powers, and initial states, the model has plenty of flexibility to fit one facility's history. The reported MAE drop from 95.8 to 20.9 kW is exactly what you'd expect from a flexible model tuned on the test set. It is not yet evidence that the \"configurable\" model transfers to other facilities or time periods.\n\nThere is also a concrete specification gap: Algorithm 1 step 11 says \"Compute Qliq(k)\" and Eq. (7) depends on it, but Qliq is never defined in the text or equations. That blocks reproduction independently of the code release. The authors mention parameter estimation as future work, which is really the missing piece for making this a generalizable tool.\n\nFor all that, the modeling approach is physically reasonable, the paper is clearly written, and the authors are transparent about the manual calibration. It deserves a serious referee, but the referee should insist on a held-out evaluation and a complete definition of Qliq before the validation claims are accepted. If those are added, this becomes a useful contribution to the data-center load-modeling literature.\n\nI'd bring it to a reading group as a useful example of how to think about thermal dynamics in grid studies, but I wouldn't cite it yet for its numbers.","headline":"Plausible RC/deadband model for simulating data-center cooling load, but the validation is in-sample only and one key term is undefined.","tokens_in":5470,"tokens_out":1544,"would_cite":false,"duration_ms":16136,"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":"A configurable RC-network model simulates hybrid data center cooling electricity demand, reducing mean absolute error from 95.80 to 20.88 kW on Marconi100 telemetry versus a constant-COP baseline.","keywords":["AI data center","cooling load simulation","hybrid cooling","RC thermal network","deadband control","coefficient of performance","Marconi100","demand flexibility"],"falsifier":"Run the published model on measured telemetry from a different hybrid-cooled AI data center without retuning, with an optimized constant-COP baseline for comparison; if the RC model's MAE does not beat the baseline by a large margin, the claim of broad configurable accuracy collapses.","tokens_in":4488,"feed_emoji":"❄️","tokens_out":4711,"duration_ms":38577,"temperature":0.7,"pith_summary":"The paper tries to establish that cooling electricity consumption in modern AI data centers, which mix air and liquid cooling, can be simulated as the output of a small thermal-dynamic system rather than as a fixed fraction of computing power. The proposed model tracks four temperatures — rack, room air, thermal mass, and liquid coolant — through a network of thermal resistances and capacitances, and switches air and liquid cooling loads on and off with deadband controllers. Validated on telemetry from the Marconi100 supercomputer, the model reduces mean absolute error from 95.80 kW to 20.88 kW and root-mean-square error from 109.79 kW to 27.27 kW relative to a constant coefficient-of-performance baseline. A sympathetic reader would care because this makes long-horizon power system studies of AI data centers feasible without fabricating cooling dynamics.","feed_headline":"Cooling load model cuts error from 96 to 21 kW","feed_subtitle":"A four-state thermal network with deadband control reproduces daily peaks and variability for grid studies.","key_machinery":"The central object is a four-state RC thermal network (thermal resistances and capacitances) with state vector (T_rack, T_air, T_mass, T_liq) representing rack, room air, building thermal mass, and liquid coolant temperatures, plus a fifth node for the liquid-heat-rejection loop. The state equations are discretized at 1-second resolution; a deadband (hysteresis) controller holds the air and liquid cooling units at their previous on/off state until the relevant temperature crosses a setpoint band. This mechanism lets cooling load stay flat while temperatures drift inside the deadbands, which is what constant-COP models cannot represent.","core_discovery":"On the paper's own terms, the central discovery is that the time-varying electricity consumed by a hybrid air/liquid data center cooling plant can be reproduced by a four-state RC thermal network plus a single heat-rejection node, driven by IT heat and ambient temperature, with simple hysteresis (deadband) controllers for the air and liquid loops. When configured for Marconi100, this model tracks measured cooling power far more closely than a constant-COP estimate, captures the observed weak correlation between IT power and cooling power, and reproduces the daily peak and variability statistics over roughly 520 daily profiles. The authors present the model as a configurable tool for generati","pith_inferences":["The reported calibration was done manually on the same Marconi100 data used for evaluation, so the accuracy numbers likely reflect an upper bound; a proper held-out test or an automatic parameter-estimation procedure would be needed to certify the model's portability to other facilities.","Because the model outputs cooling power at 1-second resolution and can be driven by forecast IT and weather traces, it could be embedded in optimal scheduling or demand-response simulations to quantify how much cooling load can be deferred without violating temperature constraints.","The deadband widths and setpoints could be turned into tunable flexibility parameters, letting grid operators ask 'how much headroom does this facility have for load shifting at 3 p.m.?' without access to proprietary building management data.","The same state-space structure could be extended with a model for server fan power or PUE variations, using the rack temperature state as a proxy for computing thermal state."],"forward_implications":["Power system planners can generate long-duration (year-scale) cooling load profiles for proposed AI data centers without equipment-level simulation, using only IT power and ambient temperature as inputs.","Cooling load flexibility estimates — the amount of load that can be shifted while respecting temperature deadbands — become computable from the same state-space model.","The model's physical interpretability (capacitances, resistances, setpoints) gives engineers a way to reason about how design choices change cooling demand.","Public availability of the implementation and sample data lets other researchers reproduce the reported errors on Marconi100 and adapt the model to other facilities.","The demonstrated correlation statistics suggest the model can reproduce the decoupling between IT load and cooling load that constant-COP models miss."],"fun_headline_variants":["Cooling model cuts AI data center error by 78%","Four-state RC model reproduces AI cooling load peaks","Hybrid cooling model error drops from 96 to 21 kW","Configurable model creates dynamic cooling profiles for grid studies"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing assumption is that parameters manually calibrated on one dataset (Marconi100) and validated on the same dataset give a fair estimate of predictive skill; if the fit has overfit this facility and time window, the reported error reductions will not transfer to other data centers.","fun_headline_variants_meta":{"raw":{"variants":["Cooling model cuts AI data center error by 78%","Four-state RC model reproduces AI cooling load peaks","Hybrid cooling model error drops from 96 to 21 kW","Configurable model creates dynamic cooling profiles for grid studies"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000938,"raw_usage":{"total_tokens":3815,"prompt_tokens":680,"completion_tokens":3135,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":424,"completion_tokens_details":{"reasoning_tokens":3067}},"tokens_in":424,"tokens_out":3135,"duration_ms":19479,"temperature":1.0,"reasoning_tokens":3067,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T16:24:49.220039+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the published model on measured telemetry from a different hybrid-cooled AI data center without retuning, with an optimized constant-COP baseline for comparison; if the RC model's MAE does not beat the baseline by a large margin, the claim of broad configurable accuracy collapses.","supporting_citations":[],"review_version":1}