PaperlayerView exact revision on arXiv

Source-linked AI summary

Modeling Liquid-Cooled Data Centers in EnergyPlus

Yun-Dam Ko, Dipashreya A. Sur, Cade Napier, Carlos D. Díaz-Marín, Rishee K. Jain

arXiv:2610.33148v1eess.SYcs.CEcs.DCcs.LGcs.SY

TL;DR

Whole-building, long-term modeling of liquid-cooled data centers remains limited, despite rising AI-related power demands. The paper develops a native EnergyPlus object for DTC cooling and validates it against Modelica. In a case study, warm-water cooling lowers PUE and facility energy relative to an air-cooled baseline while eliminating on-site water use.

  • Problem

    Existing tools have limited capability for whole-building, long-term evaluation of liquid cooling, which is needed to assess facility-level energy, water, and carbon metrics.

    Source: §1 · p. 1, §2.2 · p. 3

  • Method

    The paper develops a native EnergyPlus object that represents DTC liquid cooling within the whole-building solver and connects facility-water and rack-side loops through a CDU.

    Source: §3 · p. 4, §6 · p. 11

  • Results

    PUE fell from 1.43 to 1.25, on-site water use was eliminated, and total facility energy fell 37.9% relative to an air-cooled baseline in the case study.

    Source: §7 · p. 12

  • Takeaways & Limitations

    The object extends EnergyPlus to whole-building, long-term simulation of DTC liquid cooling for studying high-density data-center infrastructure.

    Source: §6 · p. 11, §7 · p. 12

    The steady-state object omits thermal mass and hydraulic dynamics, so it is unsuitable for transient or real-time control studies and pre-cooling strategies that depend on loop capacitance.

    Source: §6 · p. 12

Abstract

from arXiv · show

As AI workloads accelerate data center transition toward high-density racks with direct-to-chip (DTC) liquid cooling, existing simulation tools remain limited in evaluating whole-building and multi-year efficiency. To bridge this gap, we develop a native EnergyPlus object, ElectricEquipment:ITE:LiquidCooled, that models DTC liquid cooling for chilled-water and warm-water architectures as plant-loop components. It couples facility-water loops, rack-side secondary loops, coolant distribution units, and chip-junction temperatures within a single solver. Validation against an independent Modelica implementation across 21 operating conditions, including two full-year runs compared at one-minute resolution, yields loop-temperature RMSE at or below 0.021 K, well within the 0.1 K acceptance threshold. Our case study demonstrates that warm-water liquid cooling reduces annual PUE from 1.43 to 1.25, eliminates on-site water use, and cuts total facility energy by 37.9% relative to an air-cooled baseline. By embedding liquid-cooling logic into EnergyPlus, our work provides whole-building, long-term energy modeling for high-density AI infrastructure within an open-source framework.

1 Introduction

Rising data-center and AI power demands make holistic modeling increasingly important, but EnergyPlus lacks a liquid-cooled IT load object for facility-level analysis. This work develops and validates a scalable liquid-cooling object for EnergyPlus.

Motivation

  • Data-center energy use could reach 9.5–15.3% of U.S. electricity consumption by 2030, while more than half of new campuses are predicted to exceed 500 MW by 2035.The passage motivates holistic modeling to identify building strategies that lower power demand and improve efficiency and flexibility.Source: §1 · p. 1
  • AI racks can reach 30–100+ kW per rack, while traditional air-cooled racks operate at 7–10 kW, motivating liquid cooling for high-density workloads.The cited passage also reports chip heat fluxes up to 106 W/cm2.Source: §1 · p. 1

Modeling gap

  • EnergyPlus supports whole-building analysis with mature plant equipment models, but lacks the liquid-cooled IT load object needed to connect those models to liquid cooling.The passage identifies Modelica Buildings as a high-fidelity alternative, while noting its steep learning curve and limited practical use for multi-year facility-scale runs.Source: §1 · p. 1

Contribution

  • The work develops a scalable EnergyPlus liquid-cooling object and validates it against an independent Modelica implementation.The study then applies the object to a mid-sized data-center case study.Source: §1 · p. 1

2 Background

EnergyPlus data-center modeling builds on an air-cooled ITE foundation, while existing liquid-cooling tools do not readily support standard whole-building, long-term studies. The paper targets this facility-scale gap with a liquid-cooled object and focuses on DTC architectures across chilled- and warm-water operating points.

2.1 EnergyPlus for Data Centers

  • The air-cooled EnergyPlus object established an openly accessible foundation used with building-energy toolchains and later data-center modeling, control, and benchmark work.The cited follow-on areas include energy, water, and thermal dynamics across cooling configurations, predictive and reinforcement-learning control, and benchmark testbeds [22].Source: §2.1 · p. 2
  • The liquid-cooled ITE object extends EnergyPlus data-center modeling beyond the air-cooled foundation established by an ITE object for air distribution and temperature offsets.The prior object supports comparative evaluation of air-cooled topologies and efficiency standards.Source: §2.1 · p. 2

2.2 Existing tools

  • Facility-scale, long-term liquid-cooling analysis remains a gap: component and loop tools have shorter horizons or scaling constraints, while EnergyPlus integrates building and plant processes.Building-level metrics are defined over the facility boundary, and EnergyPlus is coupled to envelope, zone air, weather, on-site generation, and grid-emission inputs.Source: §2.2 · pp. 2–3

2.3 Liquid cooling architectures

  • DTC cooling places cold plates directly on CPUs and GPUs, and higher ASHRAE water-temperature classes permit warmer water while requiring chips and cold plates rated for those temperatures.The study selects W17 and W45 as representative chilled-water and warm-water extremes, respectively.Source: §2.3 · p. 3, Table 2 · p. 3

3 Method

The EnergyPlus object extends the existing IT equipment model for single-phase direct-to-chip cooling, coupling the facility-water plant loop to rack cooling and server-zone heat gains. Its calculations cover heat partition, loop temperatures, chip-junction feasibility, and efficiency metrics.

3 Method

  • The native ElectricEquipment:ITE:LiquidCooled object models single-phase direct-to-chip cooling as a demand-side plant-loop component, with a quasi-steady algebraic solution for the cooling path.Its three stages are heat partition, the primary facility-water loop, and the rack-side secondary loop and coolant distribution unit.Source: §3 · pp. 3–4

3.1 Heat partition between coolant and room air

  • The air-leakage fraction splits processor heat between liquid-captured heat delivered to the plant loop and heat added to the server-zone air, alongside inherited UPS losses.It represents non-ideal thermal capture, including bypass, radiation, and uncaptured server-fan heat.Source: §3.1 · pp. 4–5

3.2 Primary facility-water loop

  • The facility-water loop reads its operating supply temperature from the plant inlet and computes return temperature by an energy balance; zero flow redirects liquid heat to the zone.The branch writes its outlet temperature and mass flow back for the plant solver.Source: §3.2 · p. 5

3.3 CDU and rack secondary loop

  • The CDU sets rack-coolant supply to facility-water supply plus a user-specified approach temperature, while a constant-design-flow energy balance closes rack return temperature.This fixed offset does not capture approach-temperature growth as CDU duty rises toward its UA limit; the reported log-mean temperature difference does not feed back into the solution.Source: §3.3 · pp. 5–6

3.4 Chip-junction temperature

  • Chip-junction temperature is estimated with a lumped thermal resistance, treating rack units as parallel and dividing liquid heat equally among them.The model does not resolve rack flow, cold-plate convection, or die-level spreading; thermal resistance is fixed and load-independent, and exceedance hours are tracked without changing the energy solution.Source: §3.4 · p. 6

3.5 Efficiency metrics calculation

  • PUE, WUE, and EUE assess facility overhead, site water use, and recovered-heat value; EUE accounts for compute energy and reused heat and can exceed 1.0.WUE includes on-site water consumption, while modeled facility energy components exclude distribution losses downstream of the UPS.Source: §3.5 · p. 6

4 Validation

The EnergyPlus liquid-cooling object was cross-validated against an independently built Modelica implementation over annual plant runs and a 21-condition operating sweep. Sustained loop-temperature residuals stayed below the 0.1 K threshold, while the validation excludes several whole-building and plant-energy effects.

4.1 Setup

  • The annual comparison used an independently reconstructed Modelica plant, one-minute output, and matching liquid-side heat loads, with the first hour excluded from both plant comparisons.For the other 19 conditions, an ideal source enabled independent variation of design parameters, with each condition stepped through eight IT-load levels and measured after settling.Source: §4.1 · pp. 6–8

4.2 Full-year agreement

  • Across two full-year plant runs, loop-temperature RMSE was at most 0.021 K; the only error above 0.1 K was a 1.362 K warm-water maximum at isolated load steps.Table 5 summarizes the one-minute comparisons after the first hour was excluded; the transient discrepancy reflects differing control-modeling scope.Source: §4.2 · pp. 8–9, Table 5 · p. 9

4.3 Sensitivity across operating conditions

  • Across 21 operating conditions, every sustained temperature residual was at most 0.030 K, within the 0.1 K tolerance.The sweep covered liquid-cooling classes W17–W45, rack temperature rises of 10–41 K, CDU approaches of 4–10 K, half-to-twice design facility-water flow, and zero-to-full IT load.Source: §4.1 · p. 7, §4.4 · p. 9, §4.3 · p. 9

4.4 Validity and scope

  • The validation does not cover zone-air behavior, air-side heat leakage, heat-rejection equipment electricity, annual energy integrals, or multiple climates.The CDU uses a fixed user-specified approach, and the secondary rack loop is algebraic at constant mass flow.Source: §4.4 · p. 9

5 Case Study: Tampa, FL, USA

The Tampa case study compares air-cooled, chilled-water, and warm-water DTC configurations and finds substantial annual efficiency gains for warm-water cooling. It also identifies an hourly, climate-dependent trade-off between the two liquid-cooled systems.

5 Case Study: Tampa, FL, USA

  • The case study raises the air-side heat-leakage fraction from the validation model’s 0.05 to 0.20 and adds an 18 °C CRAH; warm-water needs a separate room chiller, unlike chilled-water.The chilled-water CRAH coil uses the main chilled-water loop, while 45 °C warm water cannot provide cold supply air.Source: §5 · pp. 10–11
  • Warm-water cooling lowered annual PUE from 1.43 to 1.25, WUE from 3.45 to 0, and facility energy 37.9% (23,633 to 14,668 MWh/yr) versus air cooling.Cooling overhead fell 65.1% (6,100 to 2,132 MWh/yr); the energy reduction came mainly from IT-side savings and the cooling plant.Source: §5 · p. 11
  • In the Tampa TMY3 full-year comparison, warm-water cooling raises EUE from 0.50 to 0.80, or 1.44 with 100% liquid-loop heat reuse; the latter is a theoretical upper bound, not an expected yield.Source: §5 · pp. 9, 11
  • Warm-water hourly PUE exceeded chilled-water PUE for 785 hours (9%) at minimum IT load in cool, dry weather, but the gap never exceeded 0.023.The trade-off reflects lower dry-cooler benefit at low liquid heat and the room air-cooled chiller’s lower COP; it would not occur with equal chiller efficiency.Source: §5 · p. 11
  • Tampa’s hot, humid summer limits dry-cooler free cooling, while its mild winter produces the crossover; the margin between liquid-cooling options may shift by location.Comparisons across ASHRAE climate zones are left to future work.Source: §5 · p. 11

6 Discussion and Limitations

The EnergyPlus object supports facility-level studies of DTC cooling and broader efficiency metrics, while its algebraic, steady-state formulation and the case study’s single-location scope constrain validated use.

6 Discussion and Limitations

  • The validated object embeds two-loop DTC physics in EnergyPlus, enabling whole-building liquid-cooling simulation alongside envelope, zone-air, and central-plant models.It represents facility and rack-side loops, the CDU, liquid/air heat split, and junction-to-coolant thermal resistance.Source: §6 · p. 11
  • The object supports design-point sizing, annual comparisons across climates and data-center sizes, and supervisory IT-load scheduling at hourly or coarser quasi-steady intervals.Plant-side outputs also support WUE and EUE alongside PUE, including quantitative comparison of waste-heat recovery and district-heat coupling.Source: §6 · p. 11
  • The CDU and rack hydraulics use a single design point, fixed approach temperature, constant secondary flow, and fixed junction-to-coolant resistance, limiting representation of part-load behavior and heterogeneous racks.The fixed CDU approach can over-predict at part load and bias rack-coolant and junction temperatures upward; variable secondary-pump operation is not represented.Source: §6 · pp. 11–12
  • The steady-state model omits thermal mass and hydraulic dynamics, so it cannot resolve liquid-loop transients or support transient, real-time control, or loop-capacitance-based pre-cooling studies.The reported validation covers liquid-side coupling only; the zone-air balance is outside its scope and air-side heat leakage relies on existing EnergyPlus objects.Source: §6 · p. 12
  • The case study covers one data center and one weather file, so its energy and efficiency results are specific to Tampa’s climate despite validation across varied operating conditions.The authors identify cross-climate evaluation as future work because the warm-water versus chilled-water trade-off is expected to be climate-sensitive.Source: §6 · p. 12

7 Conclusion

The paper introduces an EnergyPlus object for chilled- and warm-water DTC cooling and validates it against an independent Modelica model. The validation shows close loop-temperature agreement, and the Tampa case study reports lower PUE, no on-site water use, and reduced facility energy for warm-water cooling.

7 Conclusion

  • The new ElectricEquipment:ITE:LiquidCooled object brings chilled-water and warm-water DTC architectures into EnergyPlus.It was validated against an independently developed Modelica model with matching parameters.Source: §7 · p. 12
  • Loop-temperature RMSE stayed at or below 0.021 K across 21 operating conditions, within the 0.1 K acceptance threshold.The evaluation includes full-year chilled-water and warm-water runs compared at one-minute resolution, plus 19 swept conditions.Source: §7 · p. 12
  • In the Tampa case study, warm-water cooling reduced PUE from 1.43 to 1.25, eliminated on-site water use, and cut facility energy by 37.9% versus the air-cooled baseline.The paper presents the object as extending whole-building energy modeling for DTC cooling within an open-source framework.Source: §7 · p. 12

Declaration of Generative AI Usage

Claude (Sonnet 4.6) assisted with data-analysis and EnergyPlus object code, manuscript readability and language, and Figure 6. The authors reviewed and edited the content and take responsibility for the publication.

Declaration of Generative AI Usage

  • Claude (Sonnet 4.6) assisted with data-analysis and EnergyPlus object code, manuscript readability and language, and Figure 6; the authors reviewed the content and take responsibility for it.Source: Declaration of Generative AI Usage · pp. 12–13
Loading 2610.33148v1…