acceptodds
Under review as a conference paper at ICLR 2027

DC-XPIFM: An Explainable Physics-Informed Foundation Model for AI Data-Center Power Forecasting

Abstract

Rapid power fluctuations from AI workloads challenge the short-term load forecasts used to schedule reserves and coordinate grid operation. Supporting these decisions across heterogeneous AI data centers motivates the development of generalized forecasters that combine transferable temporal representations with facility-specific physical relationships. Time-series foundation models provide the former through pretraining, but do not explicitly enforce consistency between electrical and thermal dynamics. Introducing DC-XPIFM, which, to the best of the authors’ knowledge, is the first explainable physics-informed foundation model framework for short-term AI data-center power forecasting. The framework uses ordinary differential equations (ODEs) from a multi-node lumped thermal resistance-capacitance (RC) network to couple power consumption with GPU and memory temperatures, consistent with energy conservation and Newton’s law of cooling. These equations regularize fine-tuning, combining patterns learned during pretraining with electro-thermal constraints informed by local measurements. DC-XPIFM is instantiated on Chronos-2, TimesFM, and Moirai and compares each backbone with its unconstrained fine-tuned variants, alongside state-of-the-art (SOTA) forecasting baselines. PI-TimesFM achieves the lowest error among the evaluated methods at every tested look-back and prediction lengths. PI-TimesFM embeddings also significantly improve the five-class workload-probe accuracy relative to vanilla TimesFM. These results show that physics-informed forecasts more accurately capture and adhere to throttle-aware power changes, abrupt drops and recovery, and post-event stability during training, fine-tuning, and online inference, behaviors essential to responsive AI data-center power management. To the best of the authors' knowledge, this is also the first Bayesian inverse-PINN method for identifying electro-thermal RC parameters from operational AI data-center data, alongside an explainability analysis that connects forecasting gains to interpretable electro-thermal dynamics and operationally informative learned representations.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI data centers, time-series foundation models, physics-informed learning, short-term power forecasting, electro-thermal modeling

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_xVyMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This papers proposes DC-XPIFM, a physics-informed foundation model framework for short-term AI data-center power forecasting. The work focuses on two dataset MIT and NLR, and proposes two energy models and fine-tunes three time-series foundation models (TimesFM-2.5, Chronos-2, Moirai). Among the physics-informed versions of the foundation models and non-physics-informed baselines, PI-TimesFM has the lowest averaged MAE/MAPE. A separate Bayesian inverse PINN fits RC parameters and reports identifiability diagnostics.

Critical strengths
  • This paper addresses a very timely and practical problem on AI data center. With fast increasing AI workloads and associated power swings, it is important to be able to forecast the load accurately and understand the impact of increased load and dynamics involved.
  • The proposed physics model is simple and seems applicable to different foundation model without needing to alter the models' architectures.
  • The paper has very thorough baseline sweep.
Critical weaknesses
  • The improvements from the physics-informed versions of FM are overstated from the MAE/MAPE perspective. Looking at Figure 1, other than (d) and (e) where PI-FM follows the dynamics of the true time series, PI-FM is not able to follow the dynamics, especially in (f) where FM and PI-FM are mostly flat. Dynamics are very important feature in planning the power grid load. This suggests that PI-FM is likely not usable in practice.
  • Two separate physics models are proposed, one for MIT and one for NLR separately. This design undermines the generalization claim. Do the physics models work for an unseen dataset? Does a new model have to be proposed for AI data centers with different hardware structure?
  • The discussions of the two dataset are scattered in a few places, and are introduced in the "method" section. Structure of the paper could be improved.
Questions for authors

I think the first two points in the weakness section are important to address, as they are foundational on this paper's contribution.

4 comments
AuthorsOpenReview ↗

Question 1: improvements from the physics-informed versions of FM

We thank the reviewer for this comment. Incorporating the physics component was found to be valid and useful for modeling the hardware. As an extra loss term when a power meter is available, however, its gain is not specific to the physics. We address each part of the question in turn.

1. Controlled ablation of each term: We removed each of the four loss terms in turn, and we also used each one alone. The table gives the change in forecast error (MAE) relative to ordinary fine-tuning without physics:

TermRemovedUsed alone
Thermal balance−2.01%−1.65%
Residual rate−1.73%−1.94%
Power ramp−1.74%−1.95%
Utilisation consistency−1.66%−1.82%

The full loss gives −1.74% [−2.21, −1.27] lower error (better than the baseline).

2. Further tests to prove the improvement of incorporating Physics, especially when the power data is missing:

Without a meter, it was observed that the physics gain is even more prominent, where the physics-implied power is then the only training signal, and it lowers the error by 19–32%. The physics therefore helps most where measurements are missing.

3. The physics helps where it is used as a model of the hardware.

  • We solved the proposed ODEs without the data-driven component on NLR test runs, where temperature is computed from measured power alone, and power from measured temperature alone to show that the ODEs considered are valid and provides physically plausible results:
R2R^2 (temperature / power)Fixed 27 °C ambientConstants from the inverse PINN
Stable Diffusion0.55 / 0.770.87 / 0.85
LoRA0.73 / 0.860.90 / 0.92
  • The identified constants matter: For an oscillating load P=Pˉ+P^sin⁡ωtP=\bar P+\hat P\sin\omega t, the heat balance has the exact solution:

T(t)=Ta+RPˉ+RP^1+ω2τ2sin⁡(ωt−arctan⁡ωτ)+Ke−t/τ,τ=RC,T(t)=T_a+R\bar P+\frac{R\hat P}{\sqrt{1+\omega^2\tau^2}}\sin(\omega t-\arctan\omega\tau)+K e^{-t/\tau},\qquad \tau=RC ,

which states how strongly and how late temperature follows power. With the inverse-PINN constants (τ\tau about 1 min), it reproduces 72–107% of the response measured on MIT at periods from 5 min to 1 h, against 1–5% with the constants obtained using recursive least squares (τ\tau about 3 h). The oscillatory term in the ODE solution also explains why XPIFM captures the transients better than the vanilla model. This is further substantiated by the new metric introduced: ramp MAE, which is the error on points within 2 min of a true power ramp, where a step is a one-second change larger than twice the usual change. The results obtained are as follows:

Ramp MAE per node (W)Without fine-tuningWith PhysicsChange
TimesFM459399−13.1% [−15.8, −10.7]
Chronos-2461376−15.4% [−21.1, −9.9]
Moirai572402−29.3% [−31.5, −26.9]

The ramp MAE falls in 18 of 18 paired runs for each model and on every workload (Stable Diffusion −9% to −31%, LoRA −4% to −30%, online inference −19% to −32%). On windows that contain a power jump, the error falls by 17%, 19%, and 31%.

4. Which errors can fine-tuning and added telemetry not address? The unannounced workload change: on NLR, ordinary fine-tuning has 95 W error on steady windows but 456 W on windows with a power jump. More telemetry does not help, because temperature follows power (by about 0.2 s on NLR) and never leads it. In fact, forecasting power and temperature jointly under the physics loss worsens the power error by 4.4% and 6.4%. Moreover, the paper proved that physics influences what the model learns. As mentioned in the paper, XPIFM stays 51% closer to its pretrained weights than ordinary fine-tuning. Also, a fixed classifier on its embeddings identifies the five NLR workloads with balanced accuracy 0.893, against 0.821, and detects power events with AUROC 0.894 against 0.879 on NLR (0.568 against 0.523 on MIT).

5. Robustness: Without a meter, the physics gain is consistent, where the error falls in every workload, on every model:

WorkloadTimesFMChronos-2Moirai
Online inference−42%−54%−42%
LoRA−11%−4%−28%
Stable Diffusion−3%−6%−24%

Hence, without a meter, the gain holds at every horizon, workload, and model.

AuthorsOpenReview ↗

Question 2: Two separate physics models are proposed, one for MIT and one for NLR separately. This design undermines the generalization claim. Do the physics models work for an unseen dataset? Does a new model have to be proposed for AI data centers with different hardware structure?

We thank the reviewer for the feedback provided, which has improved the quality of the submitted manuscript. It is worth noting that the MIT Supercloud dataset and the NLR dataset do not include cooling system data, and to the best of our knowledge, these are the only two publicly available datasets showing the impact of AI workloads on power consumption. However, it is important to note that cooling loads vary slowly (minutes-to-hours). The rapid transients motivating our work (exceeding 132 kW/s at rack level) primarily originate from GPU workloads, not cooling, which is why we focus on short-term forecasting (order of minutes). This is the first work to address this challenge using our XPIFM architecture. If facility-level data were available, cooling parameters can be incorporated as an additional slow-dynamics thermal node in the RC network. Nevertheless, our RC model operates at the GPU-board level, not at the facility level.

How is DC-XPiFM instantiated at an unseen data center, and does it transfer?

MIT provides GPU and memory signals, whereas NLR provides per-GPU power and temperature plus CPU power, with no memory channel. The procedure therefore uses only what every site has: GPU temperature, power, and the GPU datasheet. However, in practice, nvidia-smi reports the same exact GPU and memory fields at every facility. To the best of our knowledge, these are the only two public datasets that show how different AI workloads affect power. The physics adapts to this, because the heat balance is modular so other parameters can be incorporated as additional nodes in the RC network, increasing the number of RC parameters. We will include this analysis and the new equations incorporating cooling in the appendix. Nevertheless, for transfer, we use the data available everywhere: GPU temperature, power, and the GPU datasheet.

Therefore,

1. The procedure is DC-XPiFM's own constraint, calibrated without a meter. That constraint is the heat balance CdT/dt=P−(T−Ta)/RC dT/dt=P-(T-T_a)/R, and setting it up at a new site means fixing its constants RR and TaT_a. With a power meter, our inverse PINN identifies them, as in the paper. Without one, we use the same equation once the temperature has settled (dT/dt=0dT/dt=0), which is accurate because a GPU responds thermally within about a second. It then becomes a straight line:

P=T−TaR.P=\frac{T-T_a}{R}.

A new site therefore follows:

  1. Read two numbers from the GPU datasheet: the idle power PidleP_{idle} and the maximum power PmaxP_{max}.
  2. Record GPU temperature, with no power meter, and take its 1st and 99th percentiles as the idle and full-load temperatures.
  3. Check that the recording covers both idle and full load. Otherwise, the procedure reports failure.
  4. Convert temperature into power with the line through the two points,

\frac{P-P_{\mathrm{idle}}}{P_{\mathrm{max}}-P_{\mathrm{idle}}}=\frac{T-T_{1\\%}}{T_{99\\%}-T_{1\\%}}

and fine-tune the forecasting model on this estimated power.

The two points fix the same constants that the inverse PINN identifies with a meter: on the NLR training workloads, they give RR = 0.045 and 0.052 K/W with TaT_a = 37.8 and 36.4 °C, against 0.040 and 0.047 K/W with 36–39 °C (on online inference, RR differs by 34%). Moreover, the training is the same, where the physics-informed loss equals training on a blend of metered and physics-implied power, and step 4 is the case in which no meter exists, so only the physics-implied power remains.

2. Does it transfer, or is it tuned to one dataset? Here, we hid the NLR power meter (H100 GPUs), set up DC-XPiFM from the datasheet and the temperature log alone, and used the hidden meter only to score the forecasts. Hence, no constant is fitted to power data. The error falls by 18.6% (TimesFM), 21.4% (Chronos-2), and 31.5% (Moirai) relative to the same forecasting model without fine-tuning. Two further results rule out dataset-specific tuning. First, constants fitted on MIT and carried over to NLR are 12.2% worse than the datasheet calibration. Second, we applied the same four steps unchanged to two more sites.

3. Each source of heterogeneity:

  • GPU hardware: on the older V100 GPUs at MIT Supercloud and OLCF Summit, the unchanged calibration is weaker (R2R^2 0.60 and 0.63). On Summit, we found the cause in step 2: the 1st percentile is about 3 K colder than the temperature at which the GPUs idle. Taking the most frequent temperature as the idle point raises R2R^2 to 0.81 and lowers the forecast error by 42–44%. This correction needs no meter, but we chose it among three candidates using metered Summit data.
  • Sampling frequency: NLR and MIT (0.1–0.2 s, as released) were resampled to one step per second (by averaging) to ensure consistency.

We have considered these aspects in the revised version of the paper and made sure that the generalization claim is not overstated.

I thank the authors for providing this detailed rebuttal. I think the authors have addressed my concern in the "Question 2", but for "Question 1", the main weakness is that the model is not able to follow the dynamics and I don't think it has been addressed. As a result, I will maintain my original assessment, as dynamics is a crucial aspect in the power system.

We thank the reviewer for their comment and for acknowledging that our response addressed Question 2. We agree that capturing power dynamics is essential. We have expanded the analysis of Figure 1, added trajectory comparisons, and evaluated ramp errors specifically to address Question 1.

We respectfully disagree with the general conclusion that the model is unable to follow dynamics. Your original review acknowledged that PI-FM follows the true dynamics in Figures 1(d) and 1(e). The expanded analysis provides further evidence across MIT Supercloud and NLR workloads:

  • Figure 1(a): Power-slope changes under high-load operation; 37.73% lower MAE than the unconstrained FM.
  • Figure 1(b): Abrupt power reduction and subsequent recovery; 57.85% lower MAE.
  • Figure 1(c): Post-event behavior with reduced forecast drift; 31.34% lower MAE.
  • Figure 1(d): Repeated oscillations during Stable Diffusion training; 30.81% lower MAE.
  • Figure 1(e): A major low-power excursion and rebound during Llama-2 LoRA fine-tuning; 29.54% lower MAE.
  • Figure 1(f): An abrupt load transition during online inference; 3.28% lower MAE.

These percentages concern the illustrated windows. The overall MIT improvements are 1.85% in MAE and 3.80% in MAPE, which we distinguish from these window-specific gains. The trajectory comparisons directly demonstrate tracking of repeated rises and falls, recovery, and workload-driven excursions.

We acknowledge your specific observation about Figure 1(f): both forecasts smooth repeated short-duration fluctuations. PI-FM captures the principal operating-level transition but does not reproduce every subsequent dip. This is a limitation, but it does not negate the dynamic tracking demonstrated in the other panels. Our claim is improved tracking of the evaluated dynamics, without guaranteeing exact reconstruction of every fluctuation, especially that this is a forecasting problem and not an estimation problem. We are forecasting several steps in the future. If it was an estimation problem (1 step in the future), we would be able to reconstruct the trajectory fully even better than the vanilla baseline FM. This analysis is also added in the revised version of the paper.

More directly, we introduced ramp MAE to quantify errors in power changes:

Ramp MAE per node (W)Without fine-tuningWith PhysicsChange
TimesFM459399−13.1% [−15.8, −10.7]
Chronos-2461376−15.4% [−21.1, −9.9]
Moirai572402−29.3% [−31.5, −26.9]

Ramp MAE decreases in all 18 paired runs for each model and on every evaluated workload: Stable Diffusion improves by 9–31%, LoRA by 4–30%, and online inference by 19–32%. On windows containing a power jump, errors decrease by 17%, 19%, and 31% for TimesFM, Chronos-2, and Moirai, respectively. These results directly address dynamic tracking beyond aggregate power-level MAE/MAPE and selected visual examples.

The method explicitly incorporates dynamic relationships during adaptation. As detailed in Section 3.3 and Appendix B, the objective includes energy-balance residuals, their step-to-step variation, and disagreement between forecast power rates and those implied by the RC equations. We also provided the analytical ODE solution and discussed its relationship to oscillation-like workloads and step-load responses.

Our previous response additionally reported that physics-informed adaptation reduces drift from pretrained weights by approximately 51%, improves NLR workload balanced accuracy from 0.821 to 0.893, and increases power-event AUROC from 0.879 to 0.894 on NLR and 0.523 to 0.568 on MIT. These representation-level results complement the direct trajectory and ramp-error evidence.

We have also added more forecast windows and comparisons among physics alone, XPIFM, and vanilla FM to make their trajectory differences explicit. The inverse-PINN analysis provides complementary physical verification through parameter identification and comparison with recursive least squares, as elaborated in the revised manuscript.

Practical deployment must be assessed against a specific operational task and timescale. However, the smoothing observed in Figure 1(f) does not establish general practical unusability. The expanded trajectories, consistent ramp-MAE reductions, and improvements on power-jump windows support our conclusion that physics-informed adaptation improves tracking of the evaluated power dynamics.

We respectfully ask that Question 1 be reassessed in light of this evidence. We also welcome any feedback that would improve the quality of the revised version of the paper.

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

This paper proposes DC-XPiFM, a physics-informed time-series foundation model framework for short-term AI data-center power forecasting. The method incorporates electro-thermal constraints derived from resistance-capacitance (RC) models into the fine-tuning objectives of TimesFM, Chronos-2, and Moirai. Evaluations on the MIT Supercloud and National Lab of the Rockies (NLR) datasets report improvements over unconstrained foundation models and conventional forecasting baselines. The paper further analyzes learned representations and employs an inverse physics-informed neural network to identify electro-thermal parameters from operational telemetry.

Critical strengths

1. Integration of physical priors into pretrained forecasters. The paper explores how electro-thermal constraints can complement transferable temporal representations, providing a potentially useful approach to incorporating physical knowledge into TSFM fine-tuning.

2. Multi-backbone evaluation and interpretability analysis. Evaluating three TSFM architectures, together with forecasting baselines, representation probes, and physical-parameter identification, offers multiple perspectives on the effects of physics-informed adaptation.

Critical weaknesses

1. The transferability of the proposed physics-informed methodology to heterogeneous AI data centers remains unclear.

The proposed framework is motivated as a generalizable forecasting approach across heterogeneous AI data centers, but its physical formulations are highly dependent on the two evaluated datasets. The MIT Supercloud setting uses an aggregate two-node GPU–memory thermal model, whereas NLR uses a per-GPU thermal model with a different residual formulation and a static prior for the reported forecasting experiments. These dataset-specific designs raise questions about how DC-XPiFM can be consistently instantiated across data centers with different GPU architectures, cooling systems, telemetry availability, and sampling frequencies. Evaluating several TSFM backbones does not establish cross-facility generalization, and the current experiments do not sufficiently demonstrate that the proposed physical constraints transfer to previously unseen AIDC configurations without substantial re-engineering.

2. The incremental benefit and necessity of the physics-informed loss are not sufficiently established.

Although the paper motivates its physical constraints through energy conservation, Newton's law of cooling, and thermal inertia, it does not adequately isolate why the proposed physics-informed objective improves forecasting beyond ordinary supervised fine-tuning or alternative forms of regularization. The larger improvements highlighted in selected operating regimes do not, by themselves, establish the general effectiveness of the proposed physical loss. More importantly, the objective combines multiple components, including thermal-balance, residual-rate, power-ramp, and utilization-consistency terms, whose individual contributions and interactions are not sufficiently isolated. Improvements in inverse physical parameter identification and embedding characteristics also do not automatically establish that these mechanisms cause the observed forecasting gains.

3. The methodology does not sufficiently model the relationship between workload scheduling and GPU power consumption.

AIDC power demand is strongly influenced by workload arrivals, job scheduling, task type, resource allocation, and execution dynamics. These factors directly affect GPU utilization and power consumption, whereas thermal states primarily describe the resulting electro-thermal response. However, the proposed methodology focuses on thermal RC constraints and measured telemetry, without adequately establishing how workload scheduling information or other operational covariates are incorporated into power forecasting. Although the MIT dataset includes utilization measurements and the paper evaluates workload information through representation probes, these experiments do not establish that the model can anticipate power changes induced by scheduling decisions. Consequently, the framework's ability to forecast abrupt workload-driven transitions, queue-drain events, and changes in resource allocation remains insufficiently characterized.

Questions for authors

1. Generalization across data centers. How can DC-XPiFM be instantiated for an unseen AI data center with different GPU hardware, cooling architecture, sampling frequency, or available sensors? Can the authors specify a reusable procedure for constructing and calibrating the physical constraints, and provide cross-facility or cross-hardware evaluation that distinguishes genuine transferability from dataset-specific model engineering?

2. Necessity and contribution of the physics-informed loss. What specific forecasting errors cannot be addressed adequately by supervised TSFM fine-tuning or by simply incorporating additional telemetry covariates? Can the authors provide controlled ablations of each physical loss component, including the thermal-balance, residual-rate, power-ramp, and utilization-consistency terms, to quantify their individual contributions and interactions? How robust are these gains across datasets, forecasting horizons, and workload-transition regimes?

3. Workload covariates and operational causality. Why does the proposed framework primarily constrain the relationship between power and thermal states rather than explicitly modeling workload scheduling and execution dynamics? Can workload-level covariates, such as job arrivals, task types, GPU allocation, request rates, or queue states where available, improve forecasting beyond the proposed physical loss? How does the method anticipate a sudden workload change before it becomes observable in the thermal measurements?

6 comments
AuthorsOpenReview ↗

Question 1. Generalization across data centers.

We thank the reviewer for the feedback provided, which has improved the quality of the submitted manuscript. We answer each one in its own comment: W1/Q1 here, followed by W2/Q2 and W3/Q3. It is worth noting that the MIT Supercloud dataset and the NLR dataset does not include cooling system data and to the best of our knowledge, these are the only two publicly available datasets showing the impact of AI workloads on power consumption. However, it is important to note that cooling loads vary slowly (minutes-to-hours). The rapid transients motivating our work (exceeding 132 kW/s at rack level) primarily originate from GPU workloads, not cooling, which is why we focus on short-term forecasting (order of minutes). This is the first work to address this challenge using our XPIFM architecture. If facility-level data were available, cooling parameters can be incorporated as an additional slow-dynamics thermal node in the RC network. Nevertheless, our RC model operates at the GPU-board level, not at the facility level.

W1/Q1. How is DC-XPiFM instantiated at an unseen data center, and does it transfer?

MIT provides GPU and memory signals, whereas NLR provides per-GPU power and temperature plus CPU power, with no memory channel. The procedure therefore uses only what every site has: GPU temperature, power, and the GPU datasheet. However, in practice, nvidia-smi reports the same exact GPU and memory fields at every facility. To the best of our knowledge, these are the only two public datasets that show how different AI workloads affect power. The physics adapts to this, because the heat balance is modular so other parameters can be incorporated as additional nodes in the RC network, increasing the number of RC parameters. We will include this analysis and the new equations incorporating cooling in the appendix. Nevertheless, for transfer, we use the data available everywhere: GPU temperature, power, and the GPU datasheet.

Therefore,

1. The procedure is DC-XPiFM's own constraint, calibrated without a meter. That constraint is the heat balance CdT/dt=P−(T−Ta)/RC dT/dt=P-(T-T_a)/R, and setting it up at a new site means fixing its constants RR and TaT_a. With a power meter, our inverse PINN identifies them, as in the paper. Without one, we use the same equation once the temperature has settled (dT/dt=0dT/dt=0), which is accurate because a GPU responds thermally within about a second. It then becomes a straight line:

P=T−TaR.P=\frac{T-T_a}{R}.

A new site therefore follows:

  1. Read two numbers from the GPU datasheet: the idle power PidleP_{idle} and the maximum power PmaxP_{max}.
  2. Record GPU temperature, with no power meter, and take its 1st and 99th percentiles as the idle and full-load temperatures.
  3. Check that the recording covers both idle and full load. Otherwise, the procedure reports failure.
  4. Convert temperature into power with the line through the two points,

\frac{P-P_{\mathrm{idle}}}{P_{\mathrm{max}}-P_{\mathrm{idle}}}=\frac{T-T_{1\\%}}{T_{99\\%}-T_{1\\%}}

and fine-tune the forecasting model on this estimated power.

The two points fix the same constants that the inverse PINN identifies with a meter: on the NLR training workloads, they give RR = 0.045 and 0.052 K/W with TaT_a = 37.8 and 36.4 °C, against 0.040 and 0.047 K/W with 36–39 °C (on online inference, RR differs by 34%). Moreover, the training is the same: as shown under W2/Q2, the physics-informed loss equals training on a blend of metered and physics-implied power, and step 4 is the case in which no meter exists, so only the physics-implied power remains.

2. Does it transfer, or is it tuned to one dataset? Here, we hid the NLR power meter (H100 GPUs), set up DC-XPiFM from the datasheet and the temperature log alone, and used the hidden meter only to score the forecasts. Hence, no constant is fitted to power data. The error falls by 18.6% (TimesFM), 21.4% (Chronos-2), and 31.5% (Moirai) relative to the same forecasting model without fine-tuning. Two further results rule out dataset-specific tuning. First, constants fitted on MIT and carried over to NLR are 12.2% worse than the datasheet calibration. Second, we applied the same four steps unchanged to two more sites.

3. Each source of heterogeneity:

  • GPU hardware: on the older V100 GPUs at MIT Supercloud and OLCF Summit, the unchanged calibration is weaker (R2R^2 0.60 and 0.63). On Summit, we found the cause in step 2: the 1st percentile is about 3 K colder than the temperature at which the GPUs idle. Taking the most frequent temperature as the idle point raises R2R^2 to 0.81 and lowers the forecast error by 42–44%. This correction needs no meter, but we chose it among three candidates using metered Summit data.
  • Sampling frequency: NLR and MIT (0.1–0.2 s, as released) were resampled to one step per second (by averaging) to ensure consistency.
AuthorsOpenReview ↗

Question 2. Necessity and contribution of the physics-informed loss.

We thank the reviewer for this comment. Incorporating the physics component was found to be valid and useful for modeling the hardware. As an extra loss term when a power meter is available, however, its gain is not specific to the physics. We address each part of the question in turn.

1. Controlled ablation of each term: We removed each of the four loss terms in turn, and we also used each one alone. The table gives the change in forecast error (MAE) relative to ordinary fine-tuning without physics:

TermRemovedUsed alone
Thermal balance−2.01%−1.65%
Residual rate−1.73%−1.94%
Power ramp−1.74%−1.95%
Utilisation consistency−1.66%−1.82%

The full loss gives −1.74% [−2.21, −1.27] lower error (better than the baseline).

2. Does the physics add value beyond ordinary fine-tuning or other regularizers? The answer depends on whether a power meter is available.

The physics loss outperforms a standard regulariser (L2-SP, which keeps the model close to its pre-trained weights): −1.74% against −0.20%. For a fixed weight λ\lambda, the loss can be rewritten as Ldata+λLphys=(1+λ) ∥P^−(P+λ1+λε)∥2+const,L_{data}+\lambda L_{phys}=(1+\lambda)\,\lVert \hat P-(P+\tfrac{\lambda}{1+\lambda}\varepsilon)\rVert^2+\text{const},

where ε\varepsilon is the error of the power implied by the heat balance. In other words, the physics-informed loss is ordinary training on a blend of metered power and physics-implied power.

Without a meter, the physics gain is even more prominent, where the physics-implied power is then the only training signal, and it lowers the error by 19–32% (see W1/Q1). The physics therefore helps most where measurements are missing.

3. The physics helps where it is used as a model of the hardware.

  • It is valid on its own: We solved the proposed ODEs without the data-driven component on NLR test runs, where temperature is computed from measured power alone, and power from measured temperature alone.
R2R^2 (temperature / power)Fixed 27 °C ambientConstants from the inverse PINN
Stable Diffusion0.55 / 0.770.87 / 0.85
LoRA0.73 / 0.860.90 / 0.92
  • The identified constants matter: For an oscillating load P=Pˉ+P^sin⁡ωtP=\bar P+\hat P\sin\omega t, the heat balance has the exact solution:

T(t)=Ta+RPˉ+RP^1+ω2τ2sin⁡(ωt−arctan⁡ωτ)+Ke−t/τ,τ=RC,T(t)=T_a+R\bar P+\frac{R\hat P}{\sqrt{1+\omega^2\tau^2}}\sin(\omega t-\arctan\omega\tau)+K e^{-t/\tau},\qquad \tau=RC ,

which states how strongly and how late temperature follows power. With the inverse-PINN constants (τ\tau about 1 min), it reproduces 72–107% of the response measured on MIT at periods from 5 min to 1 h, against 1–5% with the constants obtained using recursive least squares (τ\tau about 3 h). The oscillatory term in the ODE solution also explains why XPIFM captures the transients better than the vanilla model. This is further substantiated by the new metric introduced: ramp MAE, which is the error on points within 2 min of a true power ramp, where a step is a one-second change larger than twice the usual change. The results obtained are as follows:

Ramp MAE per node (W)Without fine-tuningWith PhysicsChange
TimesFM459399−13.1% [−15.8, −10.7]
Chronos-2461376−15.4% [−21.1, −9.9]
Moirai572402−29.3% [−31.5, −26.9]

The ramp MAE falls in 18 of 18 paired runs for each model and on every workload (Stable Diffusion −9% to −31%, LoRA −4% to −30%, online inference −19% to −32%). On windows that contain a power jump, the error falls by 17%, 19%, and 31%.

4. Which errors can fine-tuning and added telemetry not address? The unannounced workload change: on NLR, ordinary fine-tuning has 95 W error on steady windows but 456 W on windows with a power jump. More telemetry does not help, because temperature follows power (by about 0.2 s on NLR) and never leads it. In fact, forecasting power and temperature jointly under the physics loss worsens the power error by 4.4% and 6.4%. Only schedule information addresses this error (see W3/Q3). Moreover, the paper proved that physics influences what the model learns. As mentioned in the paper, XPIFM stays 51% closer to its pretrained weights than ordinary fine-tuning. Also, a fixed classifier on its embeddings identifies the five NLR workloads with balanced accuracy 0.893, against 0.821, and detects power events with AUROC 0.894 against 0.879 on NLR (0.568 against 0.523 on MIT).

5. Robustness: Without a meter, the physics gain is consistent, where the error falls in every workload, on every model:

WorkloadTimesFMChronos-2Moirai
Online inference−42%−54%−42%
LoRA−11%−4%−28%
Stable Diffusion−3%−6%−24%

Hence, with a meter, the paper's gain is reproduced at every horizon, although the placebo gains as much at each one (−1.4% to −2.6%). Without a meter, the gain holds at every horizon, workload, and model.

AuthorsOpenReview ↗

Question 3. Workload covariates and operational causality.

We thank the reviewer for this comment. We agree that this is an important lever. We therefore tested it on the three datasets, which together cover the covariates that the reviewer lists.

1. Request rates and queue state (NLR): The NLR online-inference benchmark fixes in advance when each run starts and stops, and at what request rate (200 runs). From this schedule, we build a signal B(t)B(t) that is known ahead of time: the typical power at the scheduled request rate, plus the typical idle and wind-down pattern learned from the training data, which we use in two ways. First, the model forecasts only the difference P−BP-B ("schedule residual"). Second, the model receives BB as an input.

The forecast error per node (MAE, W) is therefore:

ModelAll windowsScheduled start or stopPower jumpSteady
TimesFM, plain22316845695
TimesFM, schedule residual14418326773
Chronos-2, plain22617545999
Chronos-2, schedule residual13917126668
Chronos-2, known-future covariate14011730669

Overall, the schedule lowers the error by 31% (TimesFM) and 35% (Chronos-2), across the 24 runs conducted.

2. Abrupt transitions: On windows that contain a scheduled start or stop, the schedule residual gives little or no gain (see the table). The cause is timing, where the measured power rises about 33 s before a logged start and falls about 71 s after a logged stop. The late fall is consistent with queued requests still being served, so we cannot tell whether the early rise is a warm-up or a clock offset between the two logs. Relative to the same model without it, the error falls by 33% [−39, −27] overall and by 25% [−32, −17] on these windows.

3. Job arrivals, task types and GPU allocation: NLR has no scheduler log, so we cover this part through the other two datasets, MIT and the newly added dataset (OCLF Summit).

  • MIT Supercloud (777 labelled jobs, 26 task types; 155 test jobs). A job's task type and resource request (GPUs, nodes, CPUs, memory, time limit), all available at submission, predict its average GPU power with R2R^2 0.69 (task type alone: 0.51). Measurements alone reach 0.32 after 30 s, 0.59 after 60 s and 0.77 after 120 s, so they need one to two minutes to overtake the submission-time information. In addition, a typical start-up profile per task type lowers the error over a job's first 300 s by 62% relative to assuming that the GPU stays idle, and by 52% relative to an average profile over all jobs. The task type is the dataset's label, so a scheduler would need users to declare it.

  • OLCF Summit (15,119 test jobs). The average power of earlier finished jobs of the same name, script, user or account predicts a new job's average power at arrival with R2R^2 0.85, against 0.02 for measurements taken before arrival and 0.48 for the first 120 s after it. As an input to Chronos-2, however, this knowledge lowers the error at job arrival by only 4.6% [−6.9, −2.5].

4. How is a sudden change anticipated before it is visible in the thermal measurements? We measured which signal moves first, at the native resolution of 0.1–0.2 s (as released).

  • Across 200 scheduled starts, 199 stops and about 1,700 power steps in Stable Diffusion training, temperature reaches half its change between 0.2 s before and 1.7 s after power.
  • This agrees with the heat balance (the ODEs derived by the authors), in which temperature follows power by at most τ≈0.5\tau\approx 0.5 s on NLR and never leads it. The 0.2 s is within the 1 s averaging of the power sensor.

On steady windows, simply repeating the last measured power (59 W) still beats every schedule variant. Moreover, on Summit only 9% of abrupt power changes fall within 60 s of a job start or end, so job records cannot time changes inside a job.

The two additions answer different questions. The schedule tells the model when the load will change, lowering the error by 31–35%. The physics tells the model how power relates to temperature, thereby replacing a missing power meter (16–31% lower error; see W1/Q1) and allowing every forecast to be checked against the heat balance (see W2/Q2). This strengthens the paper even more, as it shows the complementary roles of workload scheduling and encoding physics. The revised paper includes schedule and job information as the mechanism for anticipation accordingly, along with these complementary roles.

Reviewer_77DHOpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DC-XPIFM, a framework that incorporates RC electro-thermal constraints into the fine-tuning of pretrained time-series models for AI data-center power forecasting. The framework is evaluated with TimesFM, Chronos-2, and Moirai on MIT Supercloud and NLR telemetry. The paper also investigates representation changes and physical parameter identification through a Bayesian inverse-PINN. The strongest forecasting improvements are obtained with TimesFM.

Critical strengths
  1. The paper addresses a timely and practically relevant forecasting problem.
  2. The framework combines pretrained temporal representations with interpretable physical priors.
  3. The evaluation includes multiple backbones, datasets, and complementary analyses.
Critical weaknesses
  1. Modest and backbone-dependent forecasting gains. The average forecasting improvements are relatively small, and Moirai’s mean MAE increases by 0.46% across the evaluated configurations. Forecasting results lack uncertainty estimates across repeated training runs, while NLR evidence primarily consists of selected examples without a comprehensive forecasting benchmark.

  2. Insufficient attribution of improvements to physics. The paper does not present matched controls for additional temperature supervision, generic regularization, or individual loss components. Furthermore, NLR forecasting uses a static residual, whereas the representation analysis uses a dynamic residual. This difference weakens the connection between the forecasting gains and the proposed explanation involving thermal dynamics.

  3. Inconclusive physical validation. The thermal forward diagnostics described as out-of-sample on page 9 are explicitly identified as in-sample on page 31. Moreover, the MIT thermal capacitances remain weakly identifiable, with uncertainty intervals spanning most of their admissible ranges. These limitations weaken the claims of physical parameter recovery and generalization.

Questions for authors

one concern: are the inverse-PINN parameter estimates used in forecasting fine-tuning, and which physical configuration is used for each experiment?

5 comments
AuthorsOpenReview ↗

W1. Modest and backbone-dependent forecasting gains.

We thank the reviewer for the feedback provided, which has improved the quality of the submitted manuscript. We answer each point in its own comment: W1 here, followed by W2, W3, and the question.

We agree with all three points. The new results show where the physics gives a large gain.

1. Uncertainty over repeated training runs: We repeated every run with three seeds, which gives 60 paired runs per model (20 look-back/horizon settings × 3 seeds). The table gives the change in MAE of the physics-informed model relative to ordinary fine-tuning, with 95% intervals:

Submission (one seed)Three seedsRuns improved
−1.78%−1.74% [−2.21, −1.27]49 of 60

Hence, the gain is reproducible for TimesFM. The new results with the 3 seeds for each model are added to the revised version of the paper.

2. NLR Runs: Here, we hid the NLR power meter (H100 GPUs), set up DC-XPiFM from the datasheet and the temperature log alone (procedure in our response to Reviewer 1, W1/Q1), and used the hidden meter only to score the forecasts. Hence, no constant is fitted to power data. The error falls by 18.6% (TimesFM), 21.4% (Chronos-2), and 31.5% (Moirai) relative to the same forecasting model without fine-tuning. Without a meter, the physics gain is consistent, where the error falls in every workload, on every model:

WorkloadTimesFMChronos-2Moirai
Online inference−42%−54%−42%
LoRA−11%−4%−28%
Stable Diffusion−3%−6%−24%

3. Does the physics add value beyond ordinary fine-tuning or other regularizers? The answer depends on whether a power meter is available.

The physics loss outperforms a standard regulariser (L2-SP, which keeps the model close to its pre-trained weights): −1.74% against −0.20%. For a fixed weight λ\lambda, the loss can be rewritten as

Ldata+λLphys=(1+λ) ∥P^−(P+λ1+λε)∥2+const,L_{data}+\lambda L_{phys}=(1+\lambda)\,\lVert \hat P-(P+\tfrac{\lambda}{1+\lambda}\varepsilon)\rVert^2+\text{const},

where ε\varepsilon is the error of the power implied by the heat balance. In other words, the physics-informed loss is ordinary training on a blend of metered power and physics-implied power.

Without a meter, the physics gain is even more prominent, where the physics-implied power is then the only training signal, and it lowers the error by 19–32% (see Response W1/Q1 for Reviewer XPLk). The physics therefore helps most where measurements are missing.

4. The physics helps where it is used as a model of the hardware:

  • It is valid on its own: We solved the proposed ODEs without the data-driven component on NLR test runs, where temperature is computed from measured power alone, and power from measured temperature alone.
R2R^2 (temperature / power)Fixed 27 °C ambientConstants from the inverse PINN
Stable Diffusion0.55 / 0.770.87 / 0.85
LoRA0.73 / 0.860.90 / 0.92
  • The identified constants matter: For an oscillating load P=Pˉ+P^sin⁡ωtP=\bar P+\hat P\sin\omega t, the heat balance has the exact solution:

T(t)=Ta+RPˉ+RP^1+ω2τ2sin⁡(ωt−arctan⁡ωτ)+Ke−t/τ,τ=RC,T(t)=T_a+R\bar P+\frac{R\hat P}{\sqrt{1+\omega^2\tau^2}}\sin(\omega t-\arctan\omega\tau)+K e^{-t/\tau},\qquad \tau=RC ,

which states how strongly and how late temperature follows power. With the inverse-PINN constants (τ\tau about 1 min), it reproduces 72–107% of the response measured on MIT at periods from 5 min to 1 h, against 1–5% with the constants obtained using recursive least squares (τ\tau about 3 h). The oscillatory term in the ODE solution also explains why XPIFM captures the transients better than the vanilla model. This is further substantiated by the new metric introduced: ramp MAE, which is the error on points within 2 min of a true power ramp, where a step is a one-second change larger than twice the usual change. The results obtained are as follows:

Ramp MAE per node (W)Without fine-tuningWith PhysicsChange
TimesFM459399−13.1% [−15.8, −10.7]
Chronos-2461376−15.4% [−21.1, −9.9]
Moirai572402−29.3% [−31.5, −26.9]

The ramp MAE falls in 18 of 18 paired runs for each model and on every workload (Stable Diffusion −9% to −31%, LoRA −4% to −30%, online inference −19% to −32%). On windows that contain a power jump, the error falls by 17%, 19%, and 31%. Where a meter exists, however, ordinary fine-tuning on metered power remains more accurate on the two training workloads.

4. Advantage of Encoding Physics (proof with model embeddings): The paper proved that physics influences what the model learns. As mentioned in the paper, XPIFM stays 51% closer to its pretrained weights than ordinary fine-tuning. Also, a fixed classifier on its embeddings identifies the five NLR workloads with balanced accuracy 0.893, against 0.821, and detects power events with AUROC 0.894 against 0.879 on NLR (0.568 against 0.523 on MIT).

The revised version of the paper incorporates all of these new findings.

AuthorsOpenReview ↗

W2. Insufficient attribution of improvements to physics.

We thank the reviewer for this comment. Incorporating the physics component proved valid and useful for modeling the hardware. As an extra loss term, however, when a power meter is available, its gain is not specific to the physics. We address each part of the question in turn.

1. Controlled ablation of each term: We removed each of the four loss terms in turn, and we also used each one alone. The table gives the change in forecast error (MAE) relative to ordinary fine-tuning without physics:

TermRemovedUsed alone
Thermal balance−2.01%−1.65%
Residual rate−1.73%−1.94%
Power ramp−1.74%−1.95%
Utilisation consistency−1.66%−1.82%

The full loss gives −1.74% [−2.21, −1.27] lower error (better than the baseline). Every variant gives about the same gain.

2. Does the gain come from the temperature data? If the gain came from the temperatures, it should vanish when they are wrong. We therefore ran a placebo: the same physics loss, fed temperatures from an unrelated time window (the forecast loss still sees the correct power).

TimesFM, MITMAE change vs ordinary fine-tuning
Physics loss, correct temperatures−1.74% [−2.21, −1.27]
Placebo, unrelated temperatures−2.07% [−2.64, −1.52]

The placebo gains as much at every horizon (−1.4% to −2.6%) and, on the same 30 runs, slightly more (−0.33% [−0.52, −0.19]). Hence, with a meter, the gain does not come from the temperatures. The physics matters even more where the meter is missing (point 3 and W1).

3. Generic regularisation: The physics loss outperforms a standard regulariser (L2-SP, which keeps the model close to its pre-trained weights): −1.74% against −0.20%. For a fixed weight λ\lambda, the loss can be rewritten as

Ldata+λLphys=(1+λ) ∥P^−(P+λ1+λε)∥2+const,L_{data}+\lambda L_{phys}=(1+\lambda)\,\lVert \hat P-(P+\tfrac{\lambda}{1+\lambda}\varepsilon)\rVert^2+\text{const},

where ε\varepsilon is the error of the power implied by the heat balance. In other words, the physics-informed loss is ordinary training on a blend of metered power and physics-implied power. Without a meter, the physics gain is even more prominent, where the physics-implied power is then the only training signal, and it lowers the error by 19–32% (see W1). The physics therefore helps most where measurements are missing.

4. Static versus dynamic residual (NLR): The reviewer is right that the forecast for NLR used the static residual. The static residual checks that power equals the heat carried to the coolant, P=(T−Ta)/RgaP=(T-T_a)/R_{ga}, while the dynamic residual adds the heat stored in the GPU, Cg,dT/dtC_g\\,dT/dt. We ran both forms on all five NLR workloads with identical settings, seeds and inverse-PINN constants (24 paired runs each, 120 in total). The change observed in MAE relative to ordinary fine-tuning is presented in the following table:

WorkloadStaticDynamicDynamic against static
Stable Diffusion training−1.1% [−3.6, +1.3]−0.9% [−3.8, +2.0]+0.2% [−1.1, +1.4]
LoRA training−0.3% [−0.7, +0.1]+0.2% [−0.3, +0.6]+0.45% [+0.02, +0.88]
Online inference (rate-controlled)+0.05% [−0.18, +0.35]+0.02% [−0.25, +0.43]−0.03% [−0.12, +0.10]
Online inference (finite)+0.7% [−1.2, +2.7]+1.7% [−0.7, +4.4]+1.0% [−1.1, +3.6]
Offline inference+0.8% [+0.2, +1.5]+0.6% [−0.1, +1.4]−0.2% [−0.7, +0.3]
All 120 pairs+0.03% [−0.66, +0.68]+0.3% [−0.5, +1.1]+0.3% [−0.2, +0.9]

Hence, the dynamic term does not improve workload, and on LoRA it is slightly worse. On inference, the median difference is −0.01% to −0.03% (the +1.0% mean on finite inference comes from one run). The reason is measured:

  • Heat storage is small: Cg,dT/dtC_g\\,dT/dt averages about 2% of GPU power on training and under 0.1% on inference. During training, the time constant (0.52–0.62 s) is shorter than the 1 s forecast step, whereas during inference, the fitted capacitance is at the lower end of its range, so storage is not identifiable.
  • The temperature signal is coarse: The GPU reports temperature in whole degrees, so at 0.1–0.2 s sampling dT/dtdT/dt is dominated by rounding.
  • The heat balance alone agrees: On unseen runs, power recovered from temperature has R2R^2 of 0.853 without storage and 0.846 with it (Stable Diffusion), and 0.917 and 0.911 (LoRA).

The static form retains the part of the heat balance identified by the data.

Revision: The revised paper reports all the controls listed above and uses a single configuration for both analyses: the static residual with the identified constants.

AuthorsOpenReview ↗

W3. Inconclusive physical validation.

We thank the reviewer for this comment.

1. The diagnostics on page 9 were in-sample: They were computed on the identification windows, as page 31 states, so "out-of-sample" was an error that we have corrected. We then re-scored the same parameters, with nothing refitted, on data that the identification never saw. The table gives the temperature error on MIT (MSE, K²), from the starting parameters to the fitted ones:

DataOne step aheadFour-hour rolloutHolding the last value (one step / rollout)
Identification windows (paper)41.6 → 4.83126.0 → 22.70.13 / 4.06
Validation split (unseen)17.6 → 0.9853.4 → 3.850.13 / 5.06
Test split (unseen)22.1 → 1.9265.6 → 7.640.35 / 6.44

Hence, the improvement over the starting parameters holds out-of-sample (about 11× one-step ahead and 9× in the rollout on the test split). On NLR, we measured how well the fitted heat balance holds, as the typical mismatch between its two sides relative to GPU power. For the identification windows, the mismatches are 5.8%, 6.3%, and 3.6% (Stable Diffusion, LoRA, online inference). On unseen test runs, it is 11.9%, 8.1%, and 4.1%, with nothing refitted. Before fitting, it is 58–77%.

2. Where the heat balance is valid on its own: We solved the proposed ODEs without the data-driven component on the test runs of all five NLR workloads: temperature from measured power alone, and power from measured temperature alone.

R2R^2 (temperature / power)Fixed 27 °C ambientInverse PINN (paper)Inverse PINN, weak formTemperature s.d.
Stable Diffusion0.55 / 0.770.87 / 0.850.86 / 0.879.0 K
LoRA0.73 / 0.860.90 / 0.920.88 / 0.928.9 K
Online (finite)0.60 / 0.800.82 / 0.840.80 / 0.848.4 K
Online (rate-controlled)−53.5 / −41.6−0.22 / −1.00−0.23 / −0.952.6 K
Offline−0.58 / 0.38−0.31 / 0.38−0.46 / 0.393.0 K

Hence, the heat balance alone is valid on the three workloads whose temperature varies by about 9 K, and the identified constants improve it on each.

3. Capacitances. We agree. On MIT, the posterior of CgC_g and CmC_m fills 89–91% of the admissible range in every fit, even when the other four parameters are supplied. Their medians (198 and 22.3 J/K per device) are the midpoints of the ranges (200 and 22.4 J/K), so they carry no information from the data. The cause lies in the measurement, where at one-minute sampling, heat storage is about 1% or less of the heat flow, since GPUs convert nearly 99% of electrical energy into heat through transistor switching losses, resistive losses in interconnects, and leakage currents, so heat storage is not prominent in this case. Consequently, the revised paper reports the MIT capacitances as non-identifiable.

These findings are added to the revised version of the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.