Thermal Environment Assessment for Energy-Efficient Cooling in Air-Cooled Data Centers: Measurement, CFD, and Predictive Models

Overview

The review fills a gap: thermal environment measurement, modelling approaches and their contribution to energy-efficient cooling had not been synthesised systematically. 77 studies published since 2010 were analysed for experimental measurement, CFD simulation and predictive modelling; measurement layouts, CFD software, turbulence models, mesh density, rack representation and validation errors were tabulated, and surrogate inference times were compared with CFD run times.

Findings: of the 39 experimental studies, only about 30.7% included power measurements, which limits evidence linking thermal fixes to energy savings. Temperature was measured in 100% of studies and air velocity in 66.7%, whereas power, airflow rate and pressure were monitored less. Of the 58 CFD studies, Fluent was used in 24 (41.4%) and 6SigmaDC in 12; standard k-epsilon was used in 41 (about 70.7%), with RNG, realizable, SST k-omega and LES less common. Reported CFD validation errors ranged from 0.22% to above 30% relative (mostly below 20%) and from 0.1 to 4 °C absolute; few studies checked velocity.

Mesh densities ranged from about 112 to 41,361 cells/m³, with mesh counts typically 10^5 to 10^7 for domains of tens to about 1000 m³; resolution depends more on the study's objective than on room size. Racks were modelled as porous media in about 36% of studies and as black boxes in about 31%, and porous models better capture rack pressure drop and recirculation. Of 17 predictive-model studies, 70.6% were trained on CFD-generated data; CFD runs took 5 min to 200 h, against 0.097 to 2250 s for surrogate predictions.

Limits: validation is mostly temperature-only, airflow and power are seldom measured, and most surrogates are trained on CFD output rather than measurements; the authors therefore propose a collaborative, hierarchical framework for measurement, CFD and predictive modelling.

Key points

1. Of the 39 experimental studies, only about 30.7% included power measurements, which limits evidence linking thermal fixes to energy savings.
2. Of the 58 CFD studies, Fluent was used in 24 (41.4%) and 6SigmaDC in 12; standard k-epsilon in 41 (about 70.7%).
3. Reported CFD validation errors ranged from 0.22% to above 30% relative (mostly below 20%) and from 0.1 to 4 °C absolute; few studies checked velocity.
4. Mesh densities ranged from about 112 to 41,361 cells/m³, with mesh counts typically 10^5-10^7; resolution depends more on the study's objective than on room size.
5. Racks were modelled as porous media in about 36% of studies and as black boxes in about 31%; porous models better capture rack pressure drop and recirculation.
6. Of 17 predictive-model studies, 70.6% were trained on CFD-generated data; CFD runs took 5 min to 200 h, against 0.097 to 2250 s for surrogate predictions.
7. Temperature was measured in 100% of studies and air velocity in 66.7%; power, airflow rate and pressure less often.

Using it in design

• In a design tool, set CFD accuracy expectations from this review's ranges (0.1-4 °C absolute, relative mostly below 20%) and do not limit validation scope to temperature; ask for velocity and pressure too.
• Make the porous-medium rack representation the default; a black-box rack model captures pressure drop and recirculation less well.
• Include power measurement in the measurement plan so thermal improvements can be tied to energy savings (only 30.7% of studies had it).
• If a surrogate was trained on CFD, state that it is only as accurate as the CFD and validate it against real measurements.

Who should read it

Engineers and researchers choosing data-centre thermal measurement, CFD and surrogate modelling approaches and setting accuracy expectations.

Document contents

1. 1. Introduction
2. 2. Methodology (2.1 Search Strategy, 2.2 Methodology and Objectives)
3. 3. Experimental Measurement and Analysis in Data Centers (3.1 Links Between Measurements, Thermal Metrics, and Cooling Energy Use, 3.2 Classification of Data Center Experimental Measurement Data, 3.3 Limitations of Existing Measurement Data, 3.4 Objective-Specific Measurement Parameters and Sensor Layouts)
4. 4. CFD Modeling of Thermal Flow Field Within Data Centers (4.1 Classification and Organization of CFD Modeling, 4.2 Boundary Conditions and Model Validation, 4.4 CFD-Supported Thermal Diagnosis and Applications)
5. 5. Predictive Models for Data Center Thermal Environments (5.1 Thermal Environment Prediction Models, 5.2 Factors Affecting Model Prediction, 5.3 Analysis of Model Prediction Efficiency)
6. 6. Discussion (6.1 Comparison and Assessment of Experimental Measurements, CFD, and Predictive Models, 6.2 Collaborative Framework, 6.3 Challenges and Future Directions)
7. 7. Conclusions

Related standards

• ASHRAE TC 9.9 Thermal Guidelines for Data Processing Environments

Read the document

Publication details

Authors

Haijun Tan; Shanran Wang; Wenli Liu; Pin Li

Publisher

Buildings (MDPI AG)

Year

2026

Volume

16

Issue

19

Pages

3857

Document language

English

Source type

Scientific

Access

Open licence — downloadable

DOI

10.3390/buildings16193857

Licence

CC BY 4.0 — share with attribution

Licence evidence

www.mdpi.com

Citation