Modeling Liquid-Cooled Data Centers in EnergyPlus

arXiv:2610.33148v1 [eess.SY] 23 Oct 2026

Modeling Liquid-Cooled Data Centers in EnergyPlus Yun-Dam Ko∗

Dipashreya A. Sur∗

Cade Napier∗

yundamko@stanford.edu Stanford University Stanford, California, USA

dipashre@stanford.edu Stanford University Stanford, California, USA

cnapier@stanford.edu Stanford University Stanford, California, USA

Carlos D. Díaz-Marín

Rishee K. Jain

cdiazm@stanford.edu Stanford University Stanford, California, USA

rishee.jain@stanford.edu Stanford University Stanford, California, USA

Abstract As AI workloads accelerate data center transition toward highdensity 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.

Keywords Data centers, Liquid cooling, Direct-to-chip cooling, EnergyPlus, Building energy modeling

1

Introduction

Global data center energy consumption is rising at an unprecedented rate necessitating modeling to identify sources of energy waste. Estimates indicate that data center energy use could contribute to 9.5-15.3% of U.S. electricity consumption by 2030 [33]. In parallel, the scale of new data center campuses is also increasing as over 50% are predicted to exceed 500 MW by 2035, and nearly one-third to exceed 1 GW [7]. Therefore, modeling data centers holistically can identify optimal building strategies that lower their power demands and make them more energy efficient and flexible. As data centers continue expanding to meet growing artificial intelligence (AI) demands, new liquid cooling methodologies need to be explored and deployed. Traditional air cooling uses computer room air handler (CRAH) units to circulate and cool warm air through server rack arrangements. Liquid cooling, by contrast, circulates coolant directly through components to absorb and dissipate heat. These approaches differ significantly in their suitability for ∗ Equal contribution.

supporting AI demands. Traditional air-cooled server racks operate at 7-10 kW, while AI computing racks can reach 30-100+ kW per rack [12, 23]. AI compute chips in these racks can reach heat fluxes up to 106 W/cm2 [28], making air cooled systems no longer viable given the high power density [27]. Tools capable of modeling data-center liquid cooling at the building level already exist. The Modelica Buildings library [44] supports high-fidelity simulation of liquid-cooling equipment alongside its own thermal-zone and envelope models, and can additionally be coupled to the EnergyPlus engine through Spawn-ofEnergyPlus [43]. Its use at the facility scale nevertheless remains limited in practice. Modelica has a steep learning curve despite decades of software development and focuses more on equipmentlevel components, limiting its usage for multi-year facility scale runs [26]. Therefore, it remains challenging to assess how liquid cooling interacts with the data center design. Extending the modeling capabilities at the facility level, using tools such as EnergyPlus, could allow for identification of how liquid cooling interacts with heat recovery systems, IT workloads, and grid-level renewable energy. These interactions determine facility-level efficiency metrics such as power usage effectiveness (PUE), water usage effectiveness (WUE), carbon usage effectiveness (CUE), and energy use efficiency (EUE) [14]. EnergyPlus, a dominant building energy modeling software, is well suited for data center facility-level analyses. EnergyPlus supports both auditing whole building simulation for existing data centers and thereby the planning of future ones. It contains a mature library of plant-side equipment models, including curve-based chillers, cooling towers, dry coolers, pumps, and plant-loop controls. As our work demonstrates, these EnergyPlus components can be reused to serve a liquid-cooled IT load directly rather than being re-implemented. However, EnergyPlus lacks the liquid-cooled IT load object required to interface with these equipment models. In this work, we develop a scalable liquid cooling object for EnergyPlus. We validate the object against an independent Modelica implementation to assess its accuracy in modeling liquid cooling configurations. Section 2 reviews existing approaches to liquid cooling modeling and liquid cooling architectures. Section 3 describes the key calculations applied in the liquid cooling object. Section 4 presents the validation against Modelica. Section 5 applies the object in a case study of a mid-sized data center. Section 6 and Section 7 discuss the limitations and summarize key takeaways.

Table 1: Data-center thermal simulation tools by modeling scale; this work targets the building/facility scale. Scale

Time scale

Representative tools

Role in the DC liquid-cooling workflow

Component

µs–s

ANSYS Fluent/Icepak [2], Flotherm [32], Cadence Reality DC [8], OpenFOAM [24]

Design verification of rack layout and heat-transfer surfaces; single-timestep snapshots. Cannot resolve annual operation or building coupling.

Loop

s–days

Modelica (Buildings Library) [44], Amesim [31], Simscape Fluids [36]

Dynamic response of pumps, valves, CDUs, controllers; start-up transients and control tuning. Facility-scale annual horizons are impractical.

Building/facility

1 min–years

EnergyPlus [13], TRNSYS [34], Modelica (Buildings Library) [44], Spawn-of-EnergyPlus [43]

Integrated thermal, electrical, water, and carbon flows in one engine. The only scale at which facilityboundary efficiency metrics are physically defined and where compliance and regulatory frameworks (e.g., ASHRAE 90.4 [3], LEED [41], California Title 24 [9]) operate.

Grid

5 min–decades

OpenDSS [15], GridLAB-D [25], PLEXOS [16], HOMER Pro [37]

External context: tariffs, emissions factors, demand response, portfolio dispatch. Facility thermal behavior abstracted as a constant or scheduled load.

2 Background 2.1 EnergyPlus for Data Centers

district-scale energy analysis [39]. A growing body of work builds on it, including the physics-based and data-driven modeling of data-center energy, water, and thermal dynamics across alternative cooling configurations [1, 10], the development of predictive and reinforcement learning based control strategies [6, 11, 21, 46], and the construction of open benchmark testbeds for sustainable data-center operation [22]. Building upon these air-cooled modeling foundations, in this work, we develop a liquid-cooled ITE object in EnergyPlus to extend its capabilities toward high-density, liquid-cooled data center architectures.

EnergyPlus is an open-source building energy simulation engine developed primarily by the U.S. Department of Energy (DOE), currently maintained by the National Laboratory of the Rockies, that models thermal dynamics, heating, ventilation, and air conditioning (HVAC), and energy flows across various building systems. Operating on a nodal framework under the assumption of well-mixed zone air, it traditionally simulates multi-zone energy performance at hourly or sub-hourly timesteps. While the DOE provides a suite of Commercial Prototype Building Models representing major commercial building types [38], data centers were long absent from these standardized reference models. To fill this critical gap, Sun et al. [35] introduced prototype data center energy models to enable holistic, scalable evaluations of data center cooling efficiency within a standardized building energy modeling environment. Specifically, they modified the ITE object in EnergyPlus, to feature a dedicated flow control mechanism to support the thermal dynamics of air-cooled data center spaces. This air-cooled IT equipment (ITE) object incorporates supply and return approach temperature differences to account for supply air bypass, hot exhaust recirculation, and air containment management strategies. By adjusting the effective supply air volume and return temperature based on these approach temperature differences, EnergyPlus captures realistic air distribution and temperature offsets without violating the underlying nodal heat balance framework. This integration facilitates accurate energy modeling and comparative evaluations across diverse air-cooled data center topologies, cooling strategies, and efficiency standards. The impact of that object, however, extends beyond the simulation capability it added, as it established the air-cooled data center as an openly accessible modeling foundation for subsequent research. Because it brings data-center thermal behavior into EnergyPlus, the object can be used with established building-energy toolchains (e.g., OpenStudio, DesignBuilder, Honeybee) and co-simulation interfaces, placing data-center research within whole-building and

2.2

Existing tools

Existing simulation environments that model data-center liquid cooling can be organized along two axes, spatial scale (spanning from chip to loop, facility, and grid) and temporal horizon (microseconds to decades), yielding the four scales summarized in Table 1. This work contributes primarily at the building/facility scale by enabling long-term, including annual and multi-year, whole-building simulation of DTC liquid cooling in EnergyPlus. Each tier is served by a mature tool family, but no single tier alone can express the whole-building, long-term planning under various scenarios that motivates this work. At the component level, computational fluid dynamics (CFD) tools such as ANSYS Fluent/Icepak [2], Siemens FloTHERM [32], Cadence Reality DC [8], and OpenFOAM [24] resolve chip, coldplate, and rack thermofluidics at sub-second timesteps. They are the standard for cold-plate and rack-layout design verification, but their computational cost restricts them to short transient windows and precludes annual or multi-year facility studies. At the loop level, one-dimensional thermo-fluid network tools, including Simcenter Amesim [31], Simscape Fluids [36], and, most prominently for buildings research, the Modelica ecosystem via the Lawrence Berkeley National Lab (LBNL) Buildings library [44], resolve the dynamic response of pumps, valves, heat exchangers, 2

Table 2: Facility water supply temperatures by liquid cooling class, from the 2021 ASHRAE thermal guidelines for liquid cooling [4]. The minimum water temperature for all classes is 2°C (35.6°F).

and controllers. This tier is well suited for plant-level control sequence tuning and equipment cycling, but scaling it to facility-level analysis is costly, owing to stiff hybrid differential-algebraic equation (DAE) integration, state events from equipment switching, and the translation overhead of detailed envelope models [43, 44]. As a result, annual runs of a fully coupled Modelica envelope-plusplant model are practical for targeted design and control studies, whereas load-based workflows remain faster and better supported for regulatory analysis [43]. At the building level, the tier occupied by whole-building energy simulators including EnergyPlus [13] and TRNSYS [34], thermal, electrical, water, and carbon flows are solved in a single engine at one-minute to multi-year horizons. This is the only tier at which the efficiency metrics that govern data-center design and reporting are physically defined, since each is an integral over the facility boundary rather than a loop- or component-local quantity. It is also the tier referenced by building codes and rating systems such as ASHRAE Standard 90.4 [3], LEED [41], and California Title 24 [9]. Modelica-based platforms nominally reach this tier through the Buildings library [44] and the Spawn-of-EnergyPlus coupling [43], but their application at the whole-facility scale remains largely research-grade rather than standard engineering practice due to the computational and regulatory constraints noted above. At the grid level, tools such as OpenDSS [15], GridLAB-D [25], PLEXOS [16], and HOMER Pro [37] provide the external context of tariffs, emissions factors, and demand-response signals, but they abstract each facility as a scheduled or constant load and cannot resolve internal cooling behavior. Two consequences follow. First, the tools that best resolve liquidcooling physics are confined to design-tier horizons, which is why liquid-cooled equipment is well characterized at the component and loop levels but essentially absent from operational, regulatory, and carbon-accounting workflows for whole facilities. As recently highlighted by Ganesh et al. [17], this temporal and spatial confinement forces the industry to rely on siloed evaluations that isolate single metrics like energy efficiency (e.g., PUE) while neglecting critical trade-offs within the broader energy-water-carbon nexus and dynamic thermal compliance over annual weather cycles. Second, among building-tier simulators, EnergyPlus is the tool most tightly coupled to plant, envelope, zone air, weather, on-site generation, and grid-emission inputs in a single open-source solver, and it is the reference engine for the code frameworks that govern data-center compliance. Adding a native liquid-cooling object to EnergyPlus therefore fills the specific tier-level gap identified in Table 1, rather than duplicating capability that already exists elsewhere.

2.3

Liquid Cooling Class W17 W27 W32 W40 W45 W+

Facility Water Supply Temperature °C

°F

17 27 32 40 45 >45

62.6 80.6 89.6 104 113 >113

related family, immersion cooling, instead submerges entire servers in dielectric fluid that is either pumped away as a liquid (singlephase) or boiled at the chip surface and condensed overhead (twophase). Bringing the coolant closer to the die progressively raises the allowable coolant temperature and heat-flux density, often at the cost of greater server modification. DTC has emerged as the industry standard for the high-density racks that AI workloads require, and immersion, although attractive in theoretical efficiency, is constrained by cost and the difficulty of retrofitting existing air-cooled halls [45]. The operating temperature envelope of any liquid-cooled plant is codified by the 2021 ASHRAE thermal guidelines for liquid cooling [4], which partition facility water supply temperatures into six classes (Table 2). Higher classes permit warmer water and unlock free cooling and wasteheat recovery, but require chip packages and cold plates rated for the elevated temperature. To demonstrate the capability and performance of the developed EnergyPlus object across DTC cooling applications, we select the two extremes of the ASHRAE envelope as representative operating points: W17 for conventional chilled-water designs, and W45 for warm-water designs that leverage free-cooling and heat-reuse potential. These two classes bracket the design space in which the proposed model is validated and applied.

3

Method

To model DTC liquid cooling of IT equipment within a wholebuilding energy simulation, we develop a custom EnergyPlus object, ElectricEquipment:ITE:LiquidCooled, as a native plant-loop component. Throughout our modeling process we focus on singlephase DTC cooling due to its operational maturity and widespread commercial adoption [45]. As summarized in Table 3 and shown in Figure 1, the object is derived from the existing ITE object [35] in three steps. First, we retain the fields that describe the IT load itself, such as the server zone, IT power density, load schedules, CPU power curve, and uninterruptible power supply (UPS) efficiency and loss fraction. For our case of liquid cooling, the amount of heat that goes to the zone air compared to the liquid loop is controlled by the air leakage fraction parameter. The CPU power curve is a biquadratic function of the loading fraction 𝑥 (0–1, from the loading schedule) and the rack coolant supply temperature 𝑦 = 𝑇rack,sup (°C). The existing

Liquid cooling architectures

Data-center liquid cooling is commonly classified by how close the coolant is brought to the heat source [5]. In liquid-to-room systems, chilled water feeds CRAH units or in-row coolers that condition room air before it reaches unmodified servers. In liquid-to-rack systems, rear-door heat exchangers absorb server exhaust heat at the rack boundary. In liquid-to-chip systems, more commonly called direct-to-chip (DTC), cold plates are bonded directly to central processing units (CPUs) and graphics processing units (GPUs), with residual air cooling handling lower-power components. A 3

Table 3: Fields for the air- and liquid-cooled objects in EnergyPlus 26.1. A checkmark marks applicability. Field

Air

Liq.

Fan / airflow Fan power fraction Fan airflow curves Fan power curve Design entering air temperature Supply air node Approach temperature schedules Fan end-use subcategory

✓ ✓ ✓ ✓ ✓ ✓ ✓

Server & IT load Server zone IT power density IT load schedule CPU loading schedule CPU power curve CPU end-use subcategory

✓ ✓ ✓ ✓ ✓ ✓

✓ ✓ ✓ ✓ ✓ ✓

UPS UPS end-use subcategory UPS efficiency Fraction of UPS losses to zone

✓ ✓ ✓

✓ ✓ ✓

Facility / rack coolant loop Facility Water System (FWS) supply temp. FWS flow rate FWS inlet node FWS outlet node Rack coolant flow rate Rack coolant supply temperature CDU approach temperature difference

✓ ✓ ✓ ✓ ✓ ✓ ✓

Chip thermal Junction 𝑅th (chip thermal resistance) Maximum junction temperature Air leakage fraction to room air

✓ ✓ ✓

Figure 1: Screenshot of the developed ElectricEquipment:ITE:LiquidCooled object in EnergyPlus in the IDF editor.

Whereas the air-cooled object rejects all IT heat to the zone air, the liquid-cooled object registers on the demand side of an EnergyPlus PlantLoop, the Facility Water System (FWS), so that the heat captured by the cold plates is delivered directly to the building’s chilled- or warm-water plant and handled by the plant’s supply-side equipment (e.g., chiller, cooling tower, or dry cooler). The added cooling path is deliberately quasi-steady and algebraic: at each zone timestep the object evaluates a closed-form chain of energy balances rather than iterating a discretized coolant network, keeping the object numerically robust and cheap at annual sub-hourly resolution while still resolving the temperatures that govern DTC feasibility. It is decomposed into three sequentially solved stages: first, a heat-partition step that splits the processor heat between coolant and room air; second, a primary FWS loop modeled as a real plant branch; and third, an algebraic rack-side secondary loop and CDU. Since the object is defined as a demand-side component on the FWS loop, facility-level heat rejection is automatically handled by EnergyPlus’s existing plant infrastructure. Figure 2 shows the resulting loop topology and the temperatures defined in the following subsections.

EnergyPlus air-cooled object [35, 40] uses the same curve with the server inlet air temperature as 𝑦, and the curve coefficients are user inputs. Sun et al. [35] obtained such coefficients by fitting server measurements from Sahini et al. [29] and found that CPU power is nearly independent of inlet temperature, i.e., approximately proportional to load. Although the field names retain “CPU” from the air-cooled object, the modeled load is the rack’s electric power, so the object applies equally to CPU, GPU, or other accelerator racks. All simulations in this work leave the curve unspecified, in which case the object defaults to CPU power linear in load. Second, we remove the air-side fields that have no meaning for a liquid loop, such as the server-fan power and airflow curves, the design entering-air temperature, and the supply-air node. Third, we add ten new fields that define the liquid-cooling path. The shared ITload fields are reused with no changes to obtain the instantaneous CPU electric power 𝑃cpu (W) and the UPS heat loss 𝑄 ups,loss (W), where 𝑄 ups,loss denotes the portion of UPS input power dissipated as heat inside the served zone.

3.1

Heat partition between coolant and room air

The first new field, the air-leakage fraction 𝜙 ∈ [0, 1] (dimensionless), controls how the processor (e.g., CPU, GPU, TPU) heat, which for the air-cooled object goes entirely to the zone, is split into a liquid-captured heat rate 𝑄 liq (W) and an air-leaked contribution to the zone gain 𝑄 zone (W): 𝑄 liq = (1 − 𝜙) 𝑃 cpu,

𝑄 zone = 𝜙 𝑃cpu + 𝑄 ups,loss,

(1)

where 𝑃cpu (W) is the instantaneous CPU electric power obtained from the retained IT-load fields and 𝑄 ups,loss (W) is the inherited UPS heat loss defined above. 𝑄 liq is passed to the plant loop, while 𝑄 zone , the leaked CPU heat plus the inherited UPS losses, is registered as a convective internal gain on the zone-air heat balance, where it is picked up by whatever room-air system serves the space. The airleakage fraction is the single lumped parameter through which the 4

Figure 2: Schematic of the ElectricEquipment:ITE:LiquidCooled object and its coupling to the FWS plant loop and the server zone. model represents all non-ideal thermal capture (bypass, radiation, uncaptured server-fan heat) of a real cold-plate rack.

3.2

written back for the plant solver to close the loop, so the liquid load is rejected consistently with the rest of the building model and its part-load plant controls.

Primary facility-water loop

3.3

Four new fields define the primary loop: the FWS design supply des temperature 𝑇fws,sup (°C), the FWS design volumetric flow rate 𝑉¤fws 3 -1 (m s ), and the FWS coolant inlet and outlet node names, which connect the object as a demand-side branch of the plant loop. The inlet/outlet nodes make the FWS side the only part of the model that is a true fluid network. At sizing, the design mass flow rate 𝑚¤ fws (kg s-1 ) is derived from the flow-rate field and the temperaturedependent coolant density 𝜌 fws (kg m-3 ) evaluated at the design supply temperature,   des 𝑚¤ fws = 𝜌 fws 𝑇fws,sup 𝑉¤fws, (2)

CDU and rack secondary loop

Three new fields define the CDU and secondary loop: the rack des coolant design supply temperature 𝑇rack,sup (°C), the rack coolant 3 design volumetric flow rate 𝑉¤rack (m s-1 ), and the CDU approach temperature difference Δ𝑇approach (K), a user input. The CDU, that is, the liquid-to-liquid heat exchanger separating facility water from rack coolant, is not modeled as a rated heat exchanger. Instead the user-specified approach temperature difference Δ𝑇approach (K) is applied directly to give the operating rack coolant supply temperature (𝑇rack,sup (°C)), following the approach-temperature method introduced for the air side of EnergyPlus data center models [35]:

and this flow is requested from the plant each timestep. The design supply-temperature field sets the initial condition at the start of each simulation run, but the operating FWS supply temperature 𝑇fws,sup (°C) is then read from the inlet node, that is, set by the upstream plant equipment and controls, not prescribed internally. The FWS return temperature 𝑇fws,ret (°C) follows from a first-law energy balance on the branch, 𝑄 liq 𝑇fws,ret = 𝑇fws,sup + , (3) 𝑚¤ fws 𝑐 𝑝,fws

𝑇rack,sup = 𝑇fws,sup + Δ𝑇approach .

(4)

This is a key simplification of the liquid path: it pins the rackcoolant supply a constant offset above the facility-water supply, independent of load, and therefore does not capture the growth of approach temperature as the CDU duty rises toward its UA limit. The rack coolant return temperature 𝑇rack,ret (°C) is closed by a first-law balance on the constant-design-flow secondary loop,   des 𝑚¤ rack = 𝜌 rack 𝑇rack,sup 𝑉¤rack, (5)

where 𝑐 𝑝,fws (J kg-1 K-1 ) is the coolant specific heat capacity, reevaluated at the current supply temperature. A guard handles the zero-flow case: if the loop delivers no flow, no heat can be extracted and 𝑄 liq is redirected to the zone-air gain, preventing a non-physical temperature rise. The outlet node temperature and mass flow are

𝑇rack,ret = 𝑇rack,sup +

𝑄 liq , 𝑚¤ rack 𝑐 𝑝,rack

(6)

where 𝑚¤ rack (kg s-1 ) is the rack coolant mass flow rate, 𝜌 rack (kg m-3 ) its density, and 𝑐 𝑝,rack (J kg-1 K-1 ) its specific heat. The rack coolant 5

design supply-temperature field sets the temperature at which the rack coolant properties are evaluated, which keeps the rack mass flow rate at its design value, and the initial rack coolant temperatures at the start of each simulation period. For modeling simplicity, the loop is assumed to use the same working fluid as the FWS. For reporting only, the object also computes the counter-flow log-mean temperature difference across the CDU Δ𝑇lm (K), Δ𝑇1 − Δ𝑇2 , ln(Δ𝑇1 /Δ𝑇2 ) Δ𝑇1 = 𝑇rack,ret − 𝑇fws,ret, Δ𝑇2 = 𝑇rack,sup − 𝑇fws,sup, Δ𝑇lm =

(7) (8)

Figure 3: Annual IT-load schedule: a daily step schedule with four discrete utilization levels (0.25, 0.50, 0.75, 1.00) arranged as repeating quarterly ramps. The CPU power scales linearly with the load fraction, reaching the 2,143 kW design value at full load.

with Δ𝑇1 and Δ𝑇2 (K) the local temperature differences at the two ends of the CDU. The expression reduces to the arithmetic mean in the balanced-flow limit. It does not feed back into the temperature solution. We report the log-mean temperature difference to verify if the heat exchanger assumed is realistic.

3.4

Chip-junction temperature

proposed metric designed to incentivize waste heat utilization. By accounting for both the energy directed into compute components and the reused thermal energy, EUE provides a more comprehensive assessment, potentially allowing the calculated efficiency to exceed 1.0 [14].

The final two new fields, the junction-to-coolant thermal resistance max 𝑅th (K W-1 ) and the maximum allowable junction temperature 𝑇junc (°C), govern DTC thermal feasibility. The junction-to-coolant thermal resistance combines chip packaging, thermal interface materials, and conduction path with liquid convection. It is referenced to the coolant entering the rack and absorbs liquid temperature rise along the cold plate, so its value depends on flow rate. The peak chip-junction temperature 𝑇junc (°C) is estimated with a single lumped thermal-resistance relation, 𝑄 liq 𝑅th, (9) 𝑁 units where 𝑁 units is the number of rack units on the secondary loop. The rack units are assumed to be connected in parallel, each receiving coolant at 𝑇rack,sup and an equal share of 𝑄 liq . Flow within a rack is not resolved, and any series arrangement of cold plates is lumped into 𝑅th . The loop temperatures are computed from the total flow and are independent of this assumption. Our calculation treats each rack as one equivalent thermal element between coolant and junction. It does not resolve cold-plate convection or die-level spreading, and 𝑅th is a fixed, load-independent constant. The obmax , giving a ject accumulates the hours for which 𝑇junc exceeds 𝑇junc compliance metric without altering the energy solution. 𝑇junc = 𝑇rack,sup +

3.5

PUE =

𝐸 facility , 𝐸 IT

WUE =

𝑉w , 𝐸 IT

(11)

EUE =

𝐸 compute + 𝐸 reused . 𝐸 facility

(12)

𝐸 facility = 𝐸 IT + 𝐸 UPS + 𝐸 cool + 𝐸 light,

(10)

In these definitions, 𝐸 IT is the IT electricity at the UPS output (including server fans in the air-cooled case), 𝐸 UPS the UPS conversion loss (𝜂 UPS = 0.94), 𝐸 cool the electricity of chillers, heat-rejection equipment, pumps, and air-handling fans, and 𝐸 light interior lighting. Distribution losses downstream of the UPS are not modeled. 𝑉w (L) is the total on-site water consumption, including cooling tower evaporation, drift, and blowdown as well as any adiabatic cooling or humidification water, so that WUE is expressed in L kWh-1 with 𝐸 IT in kWh. 𝐸 compute is the processor electricity and 𝐸 reused the recovered liquid-loop heat.

4 Validation 4.1 Setup

Efficiency metrics calculation

To evaluate the overall sustainability and thermal energy recovery performance of the proposed data center framework, we utilize a combination of key efficiency metrics. To fully capture both the traditional infrastructure overhead and the added value of waste heat utilization, we calculate the PUE, WUE, and EUE, as defined in Equations (10) through (12) [14, 33]. While PUE is widely used as the standard to quantify the ratio of total facility energy to IT equipment energy, it inherently treats all rejected heat as a loss and does not adequately reflect the benefits of thermal recovery. Similarly, WUE evaluates site water consumption per unit of IT energy, providing critical insight into cooling-related resource demands and potential trade-offs between energy and water efficiency [30]. To address the limitation of PUE, we incorporate EUE, a newly

To verify the correctness of the newly developed EnergyPlus object ElectricEquipment:ITE:LiquidCooled, which models a DTCcooled IT load together with its CDU and facility-water interface, we cross-validate it against the widely adopted Modelica Buildings library [44] (LBNL Buildings 13.0.0, solved in OpenModelica 1.26.3). We developed the Modelica model independently as an equationbased, component-level reconstruction of the same plant, reproducing the governing relations (IT heat balance, CDU approach, and loop energy balances) in a mature simulator solved by a different numerical method. Agreement between the two implementations therefore supports the correctness of the EnergyPlus object rather than merely its internal self-consistency. 6

Table 4: Model parameters for the two plant configurations. The upper block is common to the validation and case-study models; the lower block lists what changes between them. Spanning values are shared by both plants; split entries differ. Parameter

Chilled Water (W17)

Warm Water (W45)

Common to the validation and case-study models Data-hall floor area IT power density Peak CPU electrical power Number of rack units UPS efficiency Facility-water volumetric flow CDU approach temperature difference Facility-water supply setpoint Rack-coolant supply temperature Heat rejection (liquid loop)

557.42 m2 3 844.69 W m-2 2 143.10 kW 71 0.94 0.097 m3 s-1 2.0 K 12.0 °C 14.0 °C 2.5 MW chiller (rated reference COP 5.5) + wet tower

45.0 °C 47.0 °C 3.0 MW dry cooler

Changed from the validation setup to the case study Air-side heat-leakage fraction Zone-air heat handling CRAH cooling sourcea

0.05 (validation) → 0.20 (case study) Ideal-loads system (validation) → CRAH four-pipe fan-coil (case study) Main chilled-water loop Dedicated air-cooled chiller + room CHW loop

a Case-study only; the validation models carry no CRAH and dissipate the 5 % leakage through the ideal-loads system.

The comparison spans 21 operating conditions, covering the parameter ranges over which liquid-cooled data centers operate. The facility-water supply temperature is varied from 12 to 45 °C, reaching the defining maximum of every ASHRAE liquid-cooling class up to W45 (Table 2); the rack-coolant flow is varied to give rack temperature rises from 10.3 to 41.0 K; the CDU approach from 4 to 10 K; the facility-water flow from half to twice its design value (0.097 m3 s-1 ); and the IT load from zero to the full 2,143 kW design power. Two of the conditions are realized as complete plants and simulated over a full year: a chilled-water plant in which an electric chiller backed by a wet cooling tower supplies 12 °C water, and a warm-water plant in which a dry cooler rejects the liquid load directly to ambient air at 45 °C (Table 4). In the remaining 19 conditions the facility-water boundary condition is imposed directly, which is what allows each design parameter to be varied independently and over a range no single heat-rejection plant could deliver. All conditions share the same data hall and the same IT side, simulated using the Typical Meteorological Year 3 (TMY3) weather dataset for Tampa, FL, USA, provided by EnergyPlus. The two annual runs are driven by the IT-utilization schedule of the DOE prototype data center model [35] (Figure 3); detailed parameters are summarized in Table 4. The air-side heat-leakage fraction is set to 0.05 in both models so that the liquid heat load is identical. The 5 % air portion itself lies outside the validation scope because the Modelica model represents only the liquid side, while EnergyPlus removes this heat from the zone with an idealized air system (ZoneHVAC:IdealLoadsAirSystem) that supplies exactly the cooling required without modeling any fans, coils, or chillers. Neither model therefore contains an air-side equipment model to compare. Matching the fraction nonetheless ensures the comparison isolates the liquid cooling plant rather than confounding it with a differing heat-split assumption.

The two full-plant conditions were simulated over a full year in both tools. EnergyPlus ran at a one-minute timestep and reported every plant variable at that resolution, whereas the Modelica replica was integrated as a single continuous annual run with a variablestep solver (relative tolerance 10−6 , cold-initialized once at the start of the year) and written out on the same one-minute grid. The validation is therefore carried out at a one-minute timestep over the full year. The Modelica models use Buildings.Media. Specialized.Water.TemperatureDependentDensity as the water medium rather than the library’s default, constant-property Buildings.Media.Water, so that density follows temperature as it does in EnergyPlus. The Modelica warm-water plant is cold-initialized once at the start of the January-1 run, whereas EnergyPlus begins from its converged annual state. Its facility-water controller therefore needs about the first hour of the year to settle from this cold start, producing a start-up transient that is absent in EnergyPlus. We exclude the first hour of the year from the comparison for this reason, leaving 525,541 paired samples per variable. In the chilled-water plant the supply temperature is set directly by the chiller rather than through a controller loop, so it has no such start-up transient. The first hour is nevertheless dropped from both plants, so that the two comparisons rest on an identical set of samples. Sweeping the facility-water setpoint inside a complete plant is not viable: the dry cooler cannot hold a 32 °C supply against Tampa summer air, and driving the electric chiller to such a leaving-water temperature evaluates its performance curves far outside the range over which they were fitted, so an equipment limitation would surface as a residual of the object under test. For the remaining 19 conditions the heat-rejection equipment is therefore replaced by an ideal source in both tools, a PlantComponent: TemperatureSource on the EnergyPlus supply branch and a prescribedtemperature, prescribed-flow source in Modelica, leaving the pump, 7

the plant loop, the demand branch and the object itself unchanged. In each condition the IT load is stepped from zero to full in eight levels, and statistics are taken over the last hour of each level, once the loop has settled (480 samples per condition). One design parameter is varied at a time. For each variable we report the root-mean-square error (RMSE), the mean absolute error (MAE), and the maximum absolute error (Max) between the two simulators, with the sign convention bias = Modelica − EnergyPlus. Reporting the MAE and the maximum alongside the RMSE separates a genuinely large residual from one caused by a few isolated timesteps. Aligned with the precision guidelines [19], a 0.1 K threshold is used to validate the simulated temperatures.

4.2

Full-year agreement

The full-year agreement between the EnergyPlus object and the Modelica baseline is summarized in Table 5, computed at oneminute intervals with the first hour of the year excluded. Additionally, Figure 4 illustrates a detailed snapshot of the dynamic thermal response over a representative multi-day period for both the chilled-water and warm-water configurations. The steps on days 210 and 213 follow the IT-utilization schedule (75 % to 100 % and back). Only the return temperatures (𝑇fws,ret , 𝑇rack,ret ) change. The supply temperatures (𝑇fws,sup , 𝑇rack,sup ) are fixed by the plant setpoint and the 2 K CDU approach temperature difference, while the return temperatures rise with the extracted heat, 𝑄 liq /(𝑚¤ 𝑐 𝑝 ). The rack-coolant temperature difference (𝑇rack,ret − 𝑇rack,sup ) increases from 15.2 to 20.3 K, about four times the facility-water increase (𝑇fws,ret − 𝑇fws,sup , 3.8 to 5.0 K), because the rack flow is about one quarter of the facility-water flow. The cooling power in the lower panels, in contrast, follows the outdoor air temperature (𝑇db for the warm-water plant, 𝑇wb for the chilled-water plant) as well as the IT load (8.6 to 17.0 kW for the dry-cooler fan and 260 to 386 kW for the chiller and cooling tower), because the plant absorbs the weather to hold the supply temperature at its setpoint. Across the annual simulation, all four loop temperatures agree closely in both plants, with RMSE at or below 0.021 K and MAE at or below 0.019 K. In the chilled-water plant, the facility-water and rack-coolant supply temperatures coincide to 0.000 K (RMSE, MAE and Max), because both simulators hold the 12 °C setpoint directly rather than through a controller loop. The facility-water return temperature agrees to an RMSE of 0.004 K (Max 0.005 K), and the rack-coolant return temperature is the largest chilled-water residual at 0.021 K RMSE (MAE 0.019 K, Max 0.030 K). The warmwater plant agrees to 0.008 K RMSE and 0.003 K MAE on every loop temperature, supply and return alike, and to 0.001 K on the CDU log-mean temperature difference. The rack-coolant return temperature carries the largest sustained residual, 0.021 K RMSE in the chilled-water plant. In the warm-water plant it is no larger than the supply-side temperatures. Because the rack supply temperature matches to 0.000 K (both models pin it at the facility-water supply temperature plus the 2 K approach), this residual enters entirely through the return-side heat balance, where the rack return temperature is the rack supply temperature plus the injected heat divided by the rack-coolant heat-capacity rate. The Modelica model uses a constant water specific heat of

Figure 4: EnergyPlus and Modelica loop temperatures for the chilled-water (top) and warm-water (bottom) plants over days 209–214, with IT utilization labeled. The two traces are indistinguishable. Lower panels: EnergyPlus outdoor air temperature and cooling power, shown for context and not part of the comparison.

4,180 J kg-1 K-1 , whereas EnergyPlus takes it from its temperaturedependent water-property tables. That constant is the value for water near 47 °C, so the difference vanishes at the warm-water design point and grows toward colder supply temperatures, which is what the two plants show. The mismatch therefore scales with the injected heat, and hence with IT load, and is independent of ambient temperature, identifying it as a fluid-property evaluation difference rather than a controller artifact. It remains well within tolerance. The only residual exceeding this tolerance is the warm-water maximum, which reaches 1.362 K. Because the average error is far smaller than the maximum (MAE 0.003 K, RMSE 0.008 K), this residual comes from a few isolated timesteps rather than a sustained offset, and its origin is a difference in control-modeling scope rather than plant physics, as illustrated in Figure 5. 8

Table 5: Full-year agreement between the EnergyPlus liquid-cooled object and the Modelica baseline for the warm-water and chilled-water plants (1-min intervals, first hour excluded, 525,541 samples per variable). Variables that agree to numerical precision in both plants are omitted. Variable Facility-water supply temperature Facility-water return temperature Rack-coolant supply temperature Rack-coolant return temperature Facility-water mass flow rate (kg s-1 ) CDU log-mean temperature difference

Chilled water (W17)

Warm water (W45)

RMSE

MAE

Max

RMSE

MAE

Max

0.000 0.004 0.000 0.021 0.002 –

0.000 0.003 0.000 0.019 0.002 –

0.000 0.005 0.000 0.030 0.002 –

0.008 0.008 0.008 0.008 0.001 0.001

0.003 0.004 0.003 0.003 0.001 0.001

1.362 1.363 1.362 1.362 0.001 0.001

4.3

Sensitivity across operating conditions

Table 6 reports the remaining 19 conditions. The facility-water and rack-coolant supply temperatures agree to 0.000 K in every condition. The residual is confined to the two return temperatures and the CDU log-mean temperature difference. The largest residual over the whole sweep is 0.030 K, on the rackcoolant return temperature at the coldest facility-water supply temperature (0.019 K RMSE at 12 °C), consistent with the constant specific heat identified above. The facility-water return temperature stays at or below 0.011 K and the CDU log-mean temperature difference at or below 0.007 K, while the facility-water mass flow rate agrees to 0.023 kg s-1 or better. Every entry is within the 0.1 K tolerance.

4.4

Validity and scope

Over 21 operating conditions, covering every ASHRAE liquidcooling class from W17 to W45, rack-coolant temperature rises of 10 to 41 K, CDU approaches of 4 to 10 K, facility-water flows from half to twice design, and IT loads from zero to full, every sustained temperature residual is at or below 0.030 K. Each of these residuals has an identified cause in the Modelica reference model, either the constant specific heat it uses or the dry-cooler fan controller it resolves explicitly where EnergyPlus idealizes the plant. Neither is an error in the EnergyPlus object under test. The validation does not cover the zone-air model or the routing of air-side heat leakage, which in the production models is larger and is handled by a CRAH outside the liquid-side comparison. It also does not cover the heat-rejection plant and its electricity for chillers, pumps, and fans, or full annual energy integrals. The CDU is modeled at a fixed, user-specified approach rather than with a load-dependent effectiveness, the secondary rack loop is represented algebraically at constant mass flow, and all runs use a single climate.

Figure 5: Detailed view of the temperature residual spike during an isolated load step, illustrating the transient discrepancy between EnergyPlus and Modelica due to controlmodeling differences.

EnergyPlus reads the facility-water supply temperature from the plant node, which its solver drives to the 45 °C setpoint quasistatically within each timestep, so the delivered temperature tracks the setpoint without controller lag. The Modelica plant instead resolves the dry-cooler fan proportional–integral (PI) loop explicitly (integral time 120 s). At each midnight IT-load step the heat rejected to the loop changes abruptly, and the finite-rate controller briefly overcools the water before the supply temperature recovers exponentially, returning to within 0.1 K of the setpoint in under 7 min. These excursions occur only at the load steps: every one-minute sample whose error exceeds 0.1 K falls on one of the 31 days on which the utilization schedule changes level, and the chilled-water plant, whose setpoint is enforced directly by the chiller, shows no such transient at all. The deviation is therefore not a plant-physics disagreement but a consequence of the two tools’ modeling scope. EnergyPlus idealizes the plant as tracking its setpoint instantaneously each timestep, whereas the Modelica model captures the finite-rate response of the physical fan controller.

5

Case Study: Tampa, FL, USA

To demonstrate the capabilities of the developed EnergyPlus liquidcooling object and its evaluation through key efficiency metrics (PUE, WUE, and EUE), we conduct the case study for a data center located in Tampa, Florida, USA, a hot, humid climate case, driven by the same TMY3 weather dataset used for validation over a full annual run. To evaluate the energy performance, we compare a 9

Table 6: Agreement across the 19 swept conditions, reported for the same variables and in the same order as Table 5. Each entry is the worst value over the conditions of that group (settled samples, 480 per condition; K unless noted). Variables that agree to numerical precision in every condition are omitted, as in Table 5.

Variable

Facility-water supply temperature Facility-water return temperature Rack-coolant supply temperature Rack-coolant return temperature Facility-water mass flow rate (kg s-1 ) CDU log-mean temperature difference

FWS temperature 12 to 45 °C 9 conditions

Rack-coolant flow Δ𝑇rack 10 to 41 K 4 conditions

CDU approach 4 to 10 K 4 conditions

FWS flow 0.5–2× design 2 conditions

RMSE

Max

RMSE

Max

RMSE

Max

RMSE

Max

0.000 0.004 0.000 0.019 0.014 0.004

0.000 0.006 0.000 0.030 0.014 0.007

0.000 0.003 0.000 0.001 0.011 0.001

0.000 0.005 0.000 0.002 0.011 0.002

0.000 0.003 0.000 0.008 0.011 0.005

0.000 0.005 0.000 0.014 0.011 0.007

0.000 0.007 0.000 0.001 0.023 0.002

0.000 0.011 0.000 0.001 0.023 0.003

Figure 6: Schematic comparison of the three cooling architectures modeled in EnergyPlus for the case study. The air-cooled baseline relies on standard object types, whereas both the chilled-water and warm-water systems utilize our newly developed liquid-cooling object. total of three cooling configurations: an air-cooled baseline and two DTC liquid-cooled architectures (chilled-water and warm-water), as illustrated in Figure 6. The case-study models are the validated models of Section 4 with one addition. The validation models set the air-side heat-leakage fraction to 0.05 and remove that heat with an ideal-loads system, whereas the case study raises the fraction to 0.20 and adds a CRAH to remove the resulting room load. In practice, DTC liquid cooling may capture greater than 70% of rack heat load [42], while room

air removes the remainder, so a 0.20 air-side fraction reflects this split more realistically. The CRAH is a variable-flow four-pipe fancoil unit supplying air at 18 °C. All other parameters remain those of Table 4. All three configurations, including the air-cooled baseline, use the same UPS efficiency (0.94) and lighting power density (6.89 W/m2 ). Because the object releases the air-side fraction into the thermal zone as an ordinary internal gain, the CRAH and its chilled-water supply are added entirely within the input file, reusing EnergyPlus’s 10

existing zone heat balance, fan-coil, and plant-loop components without further development. Implementing the liquid-cooled object natively in EnergyPlus therefore places hybrid air- and liquidcooled data centers within the whole-building workflows (envelope, weather, schedules, sizing, and code-compliance modeling) in which the program is already established. The two liquid-cooled architectures serve the CRAH coil differently. In the warm-water model the 45 °C facility water is too warm to make cold supply air, so a small air-cooled chiller feeds a separate room chilled-water loop that serves the coil. In the chilled-water model the CRAH coil is placed directly on the main chilled-water loop, because the main chiller already delivers cold water, so no extra chiller is needed. As shown in Figure 7, the chilled-water and warm-water liquid cooling systems show significant advantages over the air-cooled baseline for all key metrics calculated. To summarize the trends between the air-cooled baseline and warm water loop, the annual PUE drops from 1.43 to 1.25 as the WUE goes from 3.45 to 0 and EUE increases from 0.50 to 0.80, or to 1.44 if the liquid-loop heat is fully reused (Figure 7 a-c). For reference, the air-cooled baseline PUE is close to the 2023 U.S. average PUE of approximately 1.4 [30], and both liquid-cooled cases fall within the range of approximately 1.07 to 1.33 simulated for a direct liquid-cooled facility [18]. Reductions in site-level water use follow from avoiding evaporative heat rejection [20]. The full-reuse EUE value assumes 100% reuse of the liquid-loop heat extraction (equivalent to approximately 80% of the IT input) as recoverable waste heat. In practice the usable fraction depends on the temperature at which the heat is delivered and on the requirements of the downstream application, so these values represent a theoretical upper bound for applications such as district heating rather than an expected yield. Additionally, we observe that the peak PUE drops from 1.68 to 1.48 from the air cooled baseline to the warm water loop for liquid cooling. The warm water loop achieves this sharp reduction by eliminating the chiller plant and cooling tower from the liquid loop, allowing for efficient heat dissipation through dry cooling to occur. Lastly, we observe that the total facility energy reduces by 37.9% (23,633 to 14,668 MWh/yr) and cooling overhead drops by 65.1% (6,100 to 2,132 MWh/yr), confirming the high energy efficiency gains from DTC liquid cooling. Of the 8,964 MWh/yr total reduction, 53.1 % comes from the IT side itself, as DTC cooling removes the serverfan power that the air-cooled baseline requires, 44.3 % from the cooling plant, and the remaining 2.7 % from the correspondingly smaller UPS losses. Interestingly, the warm-water system does not outperform the chilled-water system across all hours. The warm-water hourly PUE is higher than that of the chilled-water system for 785 hours of the year (9.0%), exclusively at the minimum 25% IT load and in cool, dry weather (mean wet-bulb 12.3 °C compared with 19.0 °C across all 25%-load hours), where the dry cooler’s benefit scales down with the liquid heat while the chilled-water plant’s watercooled chiller is at its most efficient. Furthermore, the air-cooled chiller serving the room load in the warm-water plant operates at a lower COP than the water-cooled chiller in the chilled-water plant, and this trade-off would not occur if both configurations shared the same chiller efficiency. The gap in those hours never exceeds 0.023 in PUE. Both directions of this trade-off are climate-driven.

Tampa’s hot, humid summer limits the free-cooling potential of the dry cooler relative to colder regions, while its mild winter is what produces the crossover hours at all, so the margin between the two liquid-cooled configurations should be expected to shift with location. Evaluating both configurations across a range of ASHRAE climate zones is left to future work.

6

Discussion and Limitations

The primary contribution of this work is a validated object for whole-building simulation of DTC liquid cooling inside EnergyPlus. The custom ElectricEquipment:ITE:LiquidCooled object embeds the distinctive physics of a two-loop DTC system inside the same solver that already handles the building envelope, zone-air balance, and central plant. It represents the facility-water primary loop, the rack-side secondary loop, the CDU that couples them, the liquid/air split of the chip heat, and the junction-to-coolant thermal resistance. Liquid-cooled data centers can therefore be simulated natively in EnergyPlus at the same whole-building level of resolution previously available only for air-cooled data centers. This makes it possible to study a broader range of cooling-system configurations with high fidelity in applications such as data-center siting and grid coordination, alongside conventional air-cooled designs within a common modeling framework. Because the object is now available inside EnergyPlus, the validated plant-level results support a broad class of building- and facility-level studies. These include design-point sizing of DTC cooling plants, annual energy comparisons of alternative plant configurations across climate zones and data-center sizes, and supervisory scheduling of IT load at hourly or coarser timescales where each interval can be treated as quasi-steady. Beyond these, the object also exposes the plant-side thermal and flow quantities needed to compute a broader set of efficiency metrics beyond the widely used PUE, including WUE and the recently proposed EUE [14, 33]. Because EUE credits reused thermal energy, this provides a common basis on which emerging technologies such as waste-heat recovery and district-heat coupling can be quantitatively compared against conventional designs. Combined with EnergyPlus’s standard plant components and time-varying grid emission factors, the object further supports sustainability analyses and can serve as a physically grounded input to upstream grid, siting, and scheduling tools. Additionally, although our current models use constant facility-water temperature setpoints, EnergyPlus natively supports dynamic temperature and load-based control strategies. Our liquid cooling module can therefore be applied for both data center planning and supervisory control of cooling operations. Despite these benefits, the proposed liquid-cooled object has limitations that bound its validated use. First, the CDU and the rack-side hydraulics are represented algebraically at a single design point. The coolant delivery temperature is set from a fixed, user-specified CDU approach temperature difference rather than a load-dependent heat exchanger. The secondary coolant flow rate is held constant, and the chip-to-coolant path is collapsed into a single fixed junction-to-coolant resistance rather than a resolved cold-plate model. Because the approach of a real CDU scales with the heat it transfers, a fixed approach over-predicts it at part load 11

Figure 7: Annual PUE, WUE, EUE, facility energy, and cooling overhead of the air-cooled baseline and the chilled-water (12 °C) and warm-water (45 °C) liquid-cooled configurations in Tampa, FL. and biases the rack-coolant and junction temperatures upward. The formulation also cannot represent heterogeneous rack populations or variable secondary-pump operation. Second, the object is steadystate and omits thermal mass and hydraulic dynamics, so it does not resolve the transient response of the liquid loop, as the loadstep comparison of Section 4 illustrates. It is therefore unsuited to transient or real-time control studies, or to pre-cooling strategies that depend on the capacitance of the loop itself. Third, the reported validation covers only the liquid-side coupling between the two models. The zone air balance lies outside the scope of this comparison, and the object’s air-side heat leakage path relies on existing EnergyPlus objects. Last but certainly not least, the case study covers a single data center in one location. The validation itself spans a wide range of operating conditions, but every run uses the same weather file, so the energy and efficiency metric results reported here are specific to Tampa’s climate. Future studies should extend the case study across climate zones, where the trade-off between warm-water and chilled-water cooling is expected to be most sensitive.

21 operating conditions: full-year chilled-water (W17) and warmwater (W45) plants compared at one-minute resolution, and 19 further conditions spanning facility-water supply temperatures from 12 to 45 °C, rack-coolant flow, CDU approach, and facility-water flow. Loop-temperature RMSE stays at or below 0.021 K throughout, well inside the 0.1 K acceptance threshold, confirming fidelity for whole-building simulation. In a case study for a mid-sized data center in Tampa, FL, USA, the warm-water loop reduced PUE from 1.43 to 1.25, eliminated on-site water use, and cut total facility energy by 37.9% relative to an air-cooled baseline. By embedding liquid-cooling logic directly in EnergyPlus, this contribution extends whole-building energy modeling practice in both academia and industry as data center infrastructure is rapidly shifting toward DTC liquid cooling systems for AI data centers, all within a single open-source framework.

7

Declaration of Interests

Data Availability Statement The source code is available at https://github.com/Urban-InformaticsLab/energyplus-liquid-cooling.

Conclusion

The authors declare that they have no conflict of interest.

Our work adds the ability to model liquid-cooled data centers in EnergyPlus through introducing a new object, ElectricEquipment: ITE:LiquidCooled, for both chilled-water and warm-water DTC cooling architectures. We validate the object against an independently developed Modelica model with matching parameters across

Declaration of Generative AI Usage Claude (Sonnet 4.6) was used to help write the data analysis and EnergyPlus object code. Claude (Sonnet 4.6) was also used to help 12

Evaluation of Different Cooling Topologies. SSRN Preprint (2026). doi:10.2139/ ssrn.7071714 [18] Mashhur Gheni, Henner Kerskes, and Konstantinos Stergiaropoulos. 2026. Operational analysis of the cooling system in a direct liquid-cooled data center: a measurement and simulation study on the impact of supply water temperature. Applied Energy 403 (2026), 127061. doi:10.1016/j.apenergy.2025.127061 [19] Simon Hodder and Boris Igor Palella. 2024. Measurement and Quantification of Thermal Environments; ISO 7726 - Ergonomics of the Thermal Environment Instruments and methods for measuring and monitoring physical quantities. REHVA Journal 61, 06 (June 2024), 18–22. https://www.rehva.eu/rehvajournal/chapter/measurement-and-quantification-of-thermal-environmentsiso-7726-ergonomics-of-the-thermal-environment-instruments-andmethods-for-measuring-and-monitoring-physical-quantities [20] Nuoa Lei, Jun Lu, Arman Shehabi, and Eric Masanet. 2025. The water use of data center workloads: A review and assessment of key determinants. Resources, Conservation and Recycling 219 (2025), 108310. doi:10.1016/j.resconrec.2025. 108310 [21] Muhammad Haiqal Bin Mahbod, Chin Boon Chng, Poh Seng Lee, and Chee Kong Chui. 2022. Energy saving evaluation of an energy efficient data center using a model-free reinforcement learning approach. Applied Energy 322 (2022), 119392. doi:10.1016/j.apenergy.2022.119392 [22] Avisek Naug, Antonio Guillen, Ricardo Luna, Vineet Gundecha, Desik Rengarajan, Sahand Ghorbanpour, Sajad Mousavi, Ashwin Ramesh Babu, Dejan Markovikj, Lekhapriya D Kashyap, and Soumyendu Sarkar. 2025. SustainDC: Benchmarking for Sustainable Data Center Control. arXiv:2408.07841 [cs.LG] https://arxiv.org/abs/2408.07841 [23] Nlyte Software. 2024. Data center rack power costs: A condensed analysis. https://www.nlyte.com/blog/data-center-rack-power-costs-a-condensedanalysis/. Accessed: 2026-10-14. [24] OpenCFD Ltd. 2026. OpenFOAM: The Open Source CFD Toolbox. https: //www.openfoam.com/ Accessed: 2026-07-23. [25] Pacific Northwest National Laboratory. 2026. GridLAB-D: Simulation Software for Power Distribution Systems. Technical Report. Pacific Northwest National Laboratory (PNNL) / U.S. Department of Energy. https://www.gridlabd.org/ Available online at https://www.gridlabd.org/. [26] Kaiying Qiu, Junlu Yang, Zhi Gao, and Fusuo Xu. 2024. A review of Modelica language in building and energy: Development, applications, and future prospect. Energy and Buildings 308 (2024), 113998. [27] Gamal Refai-Ahmed, Md Malekkul Islam, H. Kabbani, T. Cader, Husam Alissa, and Hoa Do. 2022. Holistic Understanding of Best Engineering Practice to Extend Cooling Limit of Next Generations of GPU. In 2022 IEEE 24th Electronics Packaging Technology Conference (EPTC). 890–897. doi:10.1109/EPTC56328.2022.10013227 [28] Haibo Ren, Zhenhua Quan, Jiaheng Zhao, and Yaohua Zhao. 2026. Experimental study and energy-efficiency analysis of a micro-heat-pipe-array-enhanced twophase immersion cooling system for data centre server chips. Thermal Science and Engineering Progress (2026), 104753. [29] Manasa Sahini, Uschas Chowdhury, Ashwin Siddarth, Tanmay Pradip, Dereje Agonafer, Roy Zeigham, Jeff Metcalf, and Steve Branton. 2017. Comparative study of high ambient inlet temperature effects on the performance of air vs. liquid cooled IT equipment. In 2017 16th IEEE Intersociety Conference on Thermal and Thermomechanical Phenomena in Electronic Systems (ITherm). 544–550. doi:10. 1109/ITHERM.2017.7992534 [30] Arman Shehabi, Sarah J. Smith, Alex Hubbard, Alex Newkirk, Nuoa Lei, Md Abu Bakar Siddik, Billie Holecek, Jonathan Koomey, Eric Masanet, and Dale Sartor. 2024. 2024 United States Data Center Energy Usage Report. Technical Report LBNL-2001637. Lawrence Berkeley National Laboratory. https: //escholarship.org/uc/item/32d6m0d1 [31] Siemens Digital Industries Software. 2026. Simcenter Amesim. Siemens. https: //www.siemens.com/en-us/products/simcenter/systems-simulation/amesim/ [32] Siemens Digital Industries Software. 2026. Simcenter Flotherm: Electronics Cooling CFD Simulation Software. https://www.siemens.com/en-us/products/simcenter/ fluids-thermal-simulation/flotherm/ Accessed: 2026-07-23. [33] Sarah J. Smith, Alex Hubbard, Alex Newkirk, Mohan Ganeshalingam, Billie Holecek, Dale Sartor, Michael Mills, and Arman Shehabi. 2026. United States Data Center Energy Usage Report: 2025 Update. Technical Report LBNL-2001758. Lawrence Berkeley National Laboratory. https://doi.org/10.71468/P1RP4F [34] Solar Energy Laboratory, UW-Madison, TESS, CSTB, and Transsolar. 2026. TRNSYS 18: A Transient System Simulation Program. University of Wisconsin-Madison, Solar Energy Laboratory, Madison, WI, USA. https://www.trnsys.com/ Accessed: May 7, 2026. [35] Kaiyu Sun, Na Luo, Xuan Luo, and Tianzhen Hong. 2021. Prototype energy models for data centers. Energy and Buildings 231 (2021), 110603. doi:10.1016/j. enbuild.2020.110603 [36] The MathWorks, Inc. 2026. Simscape Fluids: Model and simulate fluid systems. https://www.mathworks.com/products/simscape-fluids.html Accessed: 2026-0723. [37] UL Solutions. 2026. HOMER Pro: Microgrid Optimization Software. https: //homerenergy.com/homer-pro. https://homerenergy.com/homer-pro Available

improve the readability and language of this manuscript and to develop Figure 6. The author(s) reviewed and edited all content as needed and take full responsibility for the content of the publication.

Acknowledgments This work was supported in part by the Precourt Institute for Energy (PIE) and the Center for Integrated Facility Engineering (CIFE). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of PIE and/or CIFE. The authors acknowledge the contributions of Shreya Ramachandran to the preliminary ideation of the project.

References [1] Ablimit Aili, Wenjun Long, Zhiwei Cao, and Yonggang Wen. 2024. Radiative free cooling for energy and water saving in data centers. Applied Energy 359 (2024), 122672. doi:10.1016/j.apenergy.2024.122672 [2] ANSYS, Inc. 2026. Ansys Icepak: Electronics Cooling Simulation Software. https: //www.ansys.com/products/electronics/ansys-icepak Accessed: 2026-07-23. [3] ASHRAE. 2022. ANSI/ASHRAE Standard 90.4-2022: Energy Standard for Data Centers. Standard ANSI/ASHRAE Standard 90.4-2022. American Society of Heating, Refrigerating and Air-Conditioning Engineers, Peachtree Corners, GA. [4] ASHRAE 2024. 2021 Equipment Thermal Guidelines for Data Processing Environments: ASHRAE TC 9.9 Reference Card (fifth edition, revised and expanded ed.). ASHRAE, 180 Technology Parkway, Peachtree Corners, GA 30092. https://www.ashrae.org/file%20library/technical%20resources/bookstore/ supplemental%20files/therm-gdlns-5th-r-e-refcard.pdf From Thermal Guidelines for Data Processing Environments. [5] Mohammad Azarifar, Mehmet Arik, and Je-Young Chang. 2024. Liquid cooling of data centers: A necessity facing challenges. Applied Thermal Engineering 247 (2024), 123112. doi:10.1016/j.applthermaleng.2024.123112 [6] Marco Biemann, Philipp Andreas Gunkel, Fabian Scheller, Lizhen Huang, and Xiufeng Liu. 2023. Data Center HVAC Control Harnessing Flexibility Potential via Real-Time Pricing Cost Optimization Using Reinforcement Learning. IEEE Internet of Things Journal 10, 15 (2023), 13876–13894. doi:10.1109/JIOT.2023. 3263261 [7] Bloom Energy. 2026. Data Centers Plan to Reduce Reliance on Grid Finds Bloom Energy’s 2026 Power Report. https://www.bloomenergy.com/news/data-centersplan-to-reduce-reliance-on-grid-finds-bloom-energys-2026-power-report/ [8] Cadence Design Systems, Inc. 2026. Cadence Reality DC Digital Twin Datasheet. https://www.cadence.com/en_US/home/resources/datasheets/ datacenter-insight-platform-digital-twin-ds.html Accessed: 2026-07-23. 2025 Building Energy Ef[9] California Energy Commission. 2025. ficiency Standards: California Code of Regulations, Title 24, Part 6. Technical Report. California Energy Commission, Sacramento, CA. https://www.energy.ca.gov/programs-and-topics/programs/building-energyefficiency-standards/2025-building-energy-efficiency [10] Dong Chen, Chee-Kong Chui, and Poh Seng Lee. 2025. Adaptive physically consistent neural networks for data center thermal dynamics modeling. Applied Energy 377 (2025), 124637. doi:10.1016/j.apenergy.2024.124637 [11] Dong Chen, Chee-Kong Chui, and Poh Seng Lee. 2026. Physics informed machine learning based predictive control for intelligent operation of edge datacenters. Applied Energy 402 (2026), 126975. doi:10.1016/j.apenergy.2025.126975 [12] Xin Chen, Xiaoyang Wang, Ana Colacelli, Matt Lee, and Le Xie. 2026. Electricity demand and grid impacts of AI data centers: Challenges and prospects. Nexus 3, 3 (2026), 100162. doi:10.1016/j.ynexs.2026.100162 [13] Drury B. Crawley, Linda K. Lawrie, Frederick C. Winkelmann, W.F. Buhl, Y.Joe Huang, Curtis O. Pedersen, Richard K. Strand, Richard J. Liesen, Daniel E. Fisher, Michael J. Witte, and Jason Glazer. 2001. EnergyPlus: creating a new-generation building energy simulation program. Energy and Buildings 33, 4 (2001), 319–331. doi:10.1016/S0378-7788(00)00114-6 Special Issue: BUILDING SIMULATION’99. [14] Carlos D. Díaz-Marín and Zachary J. Berquist. 2025. Flipping the switch: carbonnegative and water-positive data centers through waste heat utilization. Energy & Environmental Science 18, 1 (2025), 8403–8413. doi:10.1039/D5EE02676H [15] Electric Power Research Institute. 2026. OpenDSS: Open Source Electric Power Distribution System Simulator. https://www.epri.com/pages/sa/opendss. https: //www.epri.com/pages/sa/opendss Electric Power Research Institute (EPRI). [16] Energy Exemplar. 2026. PLEXOS® Energy Modeling Software. https://www. energyexemplar.com/plexos Accessed: 2026-07-23. [17] Viswanathan Ganesh, Hongjun Li, Michael Maloney, and Wangda Zuo. 2026. The Energy-Water-Carbon Nexus of Liquid-Cooled Data Centers: A Comprehensive 13

[43] Michael Wetter, Kyle Benne, Hubertus Tummescheit, and Christian Winther. 2024. Spawn: coupling Modelica Buildings Library and EnergyPlus to enable new energy system and control applications. Journal of Building Performance Simulation 17, 2 (2024), 274–292. doi:10.1080/19401493.2023.2266414 [44] Michael Wetter, Wangda Zuo, Thierry S. Nouidui, and Xiufeng Pang. 2014. Modelica Buildings library. Journal of Building Performance Simulation 7, 4 (2014), 253–270. doi:10.1080/19401493.2013.765506 [45] Zhijun Wu, Guanyu Zhang, Shaoan Lu, Pengfei Leng, Yang Yu, Jun Deng, and Weidi Huang. 2025. A comprehensive review of cold plate liquid cooling technology for data centers. Chemical Engineering Science 310 (2025), 121525. doi:10.1016/j.ces.2025.121525 [46] Qingang Zhang, Muhammad Haiqal Bin Mahbod, Chin-Boon Chng, Poh-Seng Lee, and Chee-Kong Chui. 2024. Residual Physics and Post-Posed Shielding for Safe Deep Reinforcement Learning Method. IEEE Transactions on Cybernetics 54, 2 (2024), 865–876. doi:10.1109/TCYB.2022.3178084

online at https://homerenergy.com/homer-pro. [38] U.S. Department of Energy. 2013. Commercial Prototype Building Models. Technical Report. U.S. Department of Energy. https://www.energycodes.gov/ development/commercial/prototype_models [39] U.S. Department of Energy. 2014. EnergyPlus: DOE’s Open-Source State-of-theArt Whole Building Energy Simulation Engine. https://www.energy.gov/cmei/ buildings/articles/energyplus. Accessed: 2026-10-08. [40] U.S. Department of Energy. 2026. EnergyPlus™ Version 26.1.0 Documentation: Input Output Reference. U.S. Department of Energy. https://energyplus.net/assets/ nrel_custom/pdfs/pdfs_v26.1.0/InputOutputReference.pdf Build: 6f2e40d102. [41] U.S. Green Building Council. 2026. LEED Rating System. U.S. Green Building Council. https://www.usgbc.org/leed [42] Vertiv Group Corp. 2026. Transforming Colocation Facilities to Meet Demand for AI Compute. White Paper. Vertiv. https://www.vertiv.com/495e4c/globalassets/ documents/white-papers/transforming-colocation-facilities-to-meet-demandfor-ai-compute_white-paper_english.pdf

14

Related documents

• Modeling Liquid-Cooled Data Centers in EnergyPlus#1455629
• An overview of an architecture proposal for a high energy physics grid#280748
• Power-Flexible AI Data Centers: A New Paradigm for Grid-Responsive Compute#306996
• Introduction to Data Centers, Energy, and the Environment#435421
• Beyond Scalar Flexibility: From Eligible AI Workloads to Dependable Load Relief#972331