Researchers Viswanathan Ganesh, Hongjun Li, Michael Maloney, and Wangda Zuo have introduced DCOPTEST, a unified virtual testbed for data center cooling systems, described in the journal SoftwareX. The platform packages twelve heterogeneous cooling topologies, ranging from legacy air-cooled systems built around Computer Room Air Handlers to advanced Direct-to-Chip liquid cooling, into a single reconfigurable simulation environment. Its central claim is striking: what once took engineering teams several weeks of manual modeling to evaluate can now be simulated, benchmarked, and reported in minutes.

The problem the tool addresses is rooted in how cooling design has traditionally been done. Although decades of research have produced highly accurate mathematical models for individual components such as chillers, pumps, and cooling towers, the real engineering challenge today lies at the system level, where the dynamic interactions among dozens of components determine overall performance. Existing workflows remain fragmented, often requiring models to be rebuilt from the ground up for every design iteration. Traditional models tend to be hard-coded to specific boundary conditions, forcing extensive manual reconfiguration of control sequences just to test minor variations in layout or heat rejection method.

DCOPTEST sidesteps this rigidity with an elegant architectural trick. Rather than compiling a new model for each configuration, the platform encapsulates a reconfigurable hydraulic network inside a single pre-compiled Functional Mock-up Unit, or FMU, following the FMI co-simulation standard. An internal automated valve matrix, designated V1 through V12, dynamically alters the active fluid flow paths within the simulation, effectively building the chosen topology in the background without any model rebuilding. Engineers simply select one of twelve reference topologies, and the valve modulation logic routes coolant through the appropriate combination of chillers, cooling towers, economizers, pumps, and heat exchangers.

The twelve topologies systematically span the design space of modern data center cooling. Types 1 through 3 represent the current industry standard for enterprise facilities, using chilled water loops and wet cooling towers in configurations ranging from economizer-first to mechanical-only operation. Types 4 through 6 substitute closed-loop dry coolers for the wet infrastructure, a choice common in water-scarce regions despite the higher operating temperatures it demands. Types 7 through 9 incorporate Coolant Distribution Units that deliver liquid cooling directly to IT equipment over a wet primary side, while Types 10 through 12 represent the emerging generation of AI-optimized designs that exploit the high return coolant temperatures of liquid cooling to maximize free cooling with dry heat rejection.

Beneath the valve matrix lies a physics-based emulator implemented in the Modelica language using the open-source Modelica Buildings Library. The equation-based, declarative approach allows the simulator to handle the nonlinear, stiff, and hybrid dynamics that arise when hydraulic and thermal networks switch between free cooling, partial mechanical cooling, and full mechanical cooling modes. Two specialized signal exchange buses connect the physics to the outside world: a Feedback bus that maps internal physical variables to standardized telemetry, and a Control bus that lets an external controller override setpoints at discrete time steps without recompiling the model. This unified control bus abstraction means a single supervisory control algorithm can be tested across all twelve layouts without modifying controller code.

A native Python run-time environment orchestrates the entire workflow through four automated stages: selection, deployment, testing, and reporting. During deployment, the run-time environment can launch multiple topologies in parallel using multiprocessing, and during testing it synchronizes discrete control actions with the continuous-time differential algebraic equation solver via the fmpy library, injecting TMY3 weather data and IT load profiles as external disturbances. The reporting stage feeds simulated telemetry into a KPI Calculation Engine that numerically integrates energy, water, and carbon streams into standardized metrics: Power Usage Effectiveness, Water Usage Effectiveness, and Carbon Usage Effectiveness, alongside thermal compliance checks against ASHRAE guidelines.

The demonstration case study is deliberately punishing. The researchers ran all twelve topologies through a full annual simulation in ASHRAE Climate Zone 1A, the very hot and humid climate of Miami, Florida, where extreme wet-bulb and dry-bulb temperatures push heat rejection systems to their absolute limits. All twelve 8,760-hour annual simulations ran concurrently on a single Intel Core i7-13700HX workstation with 64 gigabytes of RAM, completing in roughly 14 minutes of wall-clock time. Simulated PUE and WUE values showed strong quantitative alignment with national reference figures for comparable cooling configurations, lending credibility to the underlying models.

The results reveal a stark resource trade-off at the heart of cooling design. Liquid-cooled topologies systematically outperformed their air-cooled counterparts on energy and carbon: Type 7 achieved the lowest overall PUE of 1.12 and a carbon intensity of 0.45 kgCO2e per kilowatt-hour, a 17 percent PUE improvement over the comparable air-cooled Type 2 at 1.35. The thermodynamic explanation lies in the heat capacity of the working fluids. Because air has a low volumetric heat capacity, air-cooled systems must move enormous volumes of air with fans whose power scales cubically with flow rate, incurring 630 to 650 megawatt-hours of annual fan energy. Liquid loops, by contrast, use compact pumps consuming only about 35 megawatt-hours per year, cutting total annual cooling energy in Type 7 to roughly 525 megawatt-hours, a 65.5 percent reduction compared with Type 2. The higher return fluid temperatures of liquid cooling, typically 35 to 45 degrees Celsius, also dramatically expand the free cooling envelope, allowing Type 7 to sustain 100 percent free cooling year-round even in Miami’s oppressive climate.

Water tells the other side of the story. Wet cooling tower configurations consumed between 2.12 and 2.57 liters per kilowatt-hour of IT energy in make-up water to replace evaporation, drift, and blowdown, while closed-loop dry cooler architectures eliminated evaporative losses entirely, at the price of elevated fan power and PUE values peaking at 1.43. Thermal reliability remained strong across the board: all liquid-cooled topologies achieved 100 percent compliance with ASHRAE liquid cooling classes W32, W40, and W45, and air-cooled designs maintained near-perfect allowable-class compliance, although economizer-first dry air designs dipped during ambient temperature peaks. A timestep sensitivity analysis confirmed that the default one-hour control step yields annual KPI values varying by less than 1 percent compared with finer steps. The tool does have limits: each simulation instance activates a single mutually exclusive topology, so hybrid facilities that mix air and liquid cooling in one whitespace are not yet supported, though the authors identify this as a future extension. Released under the BSD 3-Clause license with pre-compiled binaries for Windows and Linux, an automated test suite, and golden reference datasets for regression checks, DCOPTEST arrives at a moment when the industry urgently needs analytical agility, offering designers a fast, standardized bridge between conceptual design and rigorous operational evaluation as the thermal demands of AI continue their relentless climb.

Subject of Research: A unified virtual testbed for simulating and benchmarking data center cooling system topologies

Article Title: DCOPTEST: A novel unified virtual testbed for data center cooling systems

Article References: Ganesh, V., Li, H., Maloney, M., & Zuo, W. (2026). DCOPTEST: A novel unified virtual testbed for data center cooling systems. SoftwareX, 36, Article 103073. https://doi.org/10.1016/j.softx.2026.103073

Image Credits: AI Generated

DOI: 10.1016/j.softx.2026.103073

Keywords: data centers, cooling systems, liquid cooling, DCOPTEST, Modelica, PUE, WUE, CUE, ASHRAE, AI infrastructure, simulation, open-source software