Register or Submit a manuscript.
Quick Link
Recent Progress in Science and Engineering

(ISSN 3067-4573)

Recent Progress in Science and Engineering (ISSN 3067-4573) is an international peer-reviewed Open-Access journal published quarterly online by LIDSEN Publishing Inc. It aims to provide an advanced knowledge platform for Science and Engineering researchers, to share the recent advances on research, innovations and development in their field.

The journal covers a wide range of subfields of Science and Engineering, including but not limited to Chemistry, Physics, Biology, Geography, Earth Science, Pharmaceutical Science, Environmental Science, Mathematical and Statistical Science, Humanity and Social Science; Civil, Chemical, Electrical, Mechanical, Computer, Biological, Agricultural, Aerospace, Systems Engineering. Articles of interdisciplinary nature are also particularly welcome.

The journal publishes all types of articles in English. There is no restriction on the length of the papers. We encourage authors to be concise but present their results in as much detail as necessary.

 
 
Indexing: 
Publication Speed (median values for papers published in 2025): Submission to First Decision: 6.6 weeks; Submission to Acceptance: 9.2 weeks; Acceptance to Publication: 6.0 days (1-2 days of FREE language polishing included)
Free Publication in 2026
Current Issue: 2026  Archive: 2025
Open Access Review

Immersion Cooling Technologies in High-Density Computing: Review and Future Perspectives

Ghassan Adnan Abd , Ahmed J. Shkarah , Mustafa M. Mansour *

  1. Department of Mechanical Engineering, College of Engineering, University of Thi-Qar, Thi-Qar 64001, Iraq

* Correspondence: Mustafa M. Mansour

Academic Editor: Orlando Corigliano

Received: July 20, 2026 | Accepted: October 12, 2026 | Published: October 26, 2026

Recent Prog Sci Eng 2026, Volume 2, Issue 4, doi:10.21926/rpse.2604024

Recommended citation: Abd GA, Shkarah AJ, Mansour MM. Immersion Cooling Technologies in High-Density Computing: Review and Future Perspectives. Recent Prog Sci Eng 2026; 2(4): 024; doi:10.21926/rpse.2604024.

© 2026 by the authors. This is an open access article distributed under the conditions of the Creative Commons by Attribution License, which permits unrestricted use, distribution, and reproduction in any medium or format, provided the original work is correctly cited.

Abstract

Air cooling is being pushed to its limits in data centres, with racks now pushing tens of kilowatts and localised areas of the accelerator reaching at 100 W cm-2. The architecture most frequently suggested to replace it is immersion cooling, where electronics are in direct contact with a dielectric liquid. Still, the literature backing this replacement is very disjointed—results are reported at varying spatial scales, reported in different system boundaries, and with incomplete property and hydraulic data, making direct comparison risky. This review does not aim to be yet another descriptive survey, but to tackle this fragmentation. 90 studies published from 2015 to 2026 were identified via a documented search protocol, screened for explicit criteria and analysed within five levels of chip, package, server, tank, and facility scales. The comparative performance of single-phase (SPIC) and two-phase (TPIC) immersion cooling is systematically compared using twelve engineering criteria and the quantitative data extracted are integrated into thirteen comparison tables between coolant properties, heat-transfer coefficients, thermal resistances, pressure drops, enhancement penalties, energy metrics with explicit boundary tags, cost and environmental indicators, and long-duration test evidence. Three findings emerge. First, the reported performance envelopes of SPIC and TPIC overlap significantly more than is suggested by their mechanisms: Forced SPIC has achieved heat-transfer coefficients of 0.4-2 kW m-2 K-1, and jet or locally actuated SPIC has achieved values between 2 and 15.3 kW m-2 K-1, which spans much of the range of heat-transfer coefficient for plain-surface nucleate boiling from 1.5 to 12 kW m-2 K-1; thus, the choice of architecture depends less on the heat-transfer coefficient achieved, and more on the critical heat flux margin, vapour management and complexity of operation. Second, apparent contradictions in the literature can be traced to five identifiable sources: declared system boundaries were different, unreported coolant properties, bulk flow rate as a substitute for local availability of coolant, mismatched scales, and steady uniform heating as a substitute for realistic workloads, and once these issues are controlled, the disagreements are largely eliminated. Third, the evidence base is very unbalanced: what is known about thermal performance—measured from the chip to the server—is well developed, but multi-year cooling ageing, material compatibility, and facility-scale demonstration and standardised test protocols are relatively limited or non-existent—this is the current limiting constraint for deployment, not thermal performance. On this basis, we suggest a prioritised research agenda and minimum reporting requirements for future immersion-cooling studies.

Keywords

Immersion cooling; two-phase immersion cooling; high-density computing; AI data centres; dielectric coolants; critical heat flux; thermal-hydraulic optimization; power usage effectiveness; long-term reliability

1. Introduction

1.1 Thermal Drivers in High-Density and AI Computing

High-density computing focuses electrical power at chip, server, rack, and facility levels. The major thermodynamic factors limiting IT are local heat flux and its spatial variation, rather than the average facility power [1,2].

Air cooling remains mature, and higher density increases the required thermal conductivity, volumetric heat capacity, and heat-transfer area, increasing electrical and physical infrastructure needs [3,4].

A typical current enterprise processor is 0.5-1.5 W cm-2 on a package-average basis, while an AI/HPC accelerator is 3-10 W cm-2 with higher local hotspots, and AI racks are typically 40-130 kW. With air-side h limited to about 250 W m-2 K-1, these loads increasingly require liquid cooling [5,6,7].

1.2 Immersion Cooling Within the Liquid-Cooling Landscape

Indirect liquid cooling is done with cold plates or heat exchangers without touching the electronics. Immersion cooling involves submerging the electronics in an electrically insulating liquid, thus eliminating the air-side resistance and reducing the heat transfer path [4,5,8].

SPIC uses a liquid coolant and depends upon natural or forced convection, while TPIC relies upon boiling and condensation where latent heat plays a role in heat transfer [1,9]. Figure 1 is a comparison of both architectures with air and direct-to-chip cooling.

Click to view original image

Figure 1 High-density cooling architectures and series thermal-resistance paths, (a) air cooling, (b) direct-to-chip cold-plate cooling, (c) single-phase immersion cooling (SPIC) and (d) two-phase immersion cooling (TPIC).

The upper row depicts the heat-transfer path from the die to the facility heat-rejection equipment; the lower row breaks down the same path into series resistances, with the width of the boxes representing the approximate proportion of the total resistance of the junction-to-ambient path. Immersion architectures eliminate air-side convective resistance and the associated air-handling infrastructure, but shift the performance-limiting role to coolant distribution in SPIC and the boiling stability, vapour transport and condensation in TPIC.

Immersion removes the dominant air-side resistance, while the limiting step becomes coolant distribution and local convection for SPIC, and nucleation, vapour removal, and condensation for TPIC. Various resistance chains are why cross-architecture comparisons need to be based on the same criteria and declared scales.

1.3 What Previous Reviews Have Established, and What They Have Not

Previous reviews have defined the general role of immersion cooling, but most are descriptive. Typical weaknesses include a lack of a reproducible search protocol, a lack of a criterion-based comparison of SPIC/TPIC, lack of scale-resolved interpretation, lack of boundary-tagged energy metrics, contradiction analysis and lack of long-term deployment evidence [4,6,10,11,12] (Table 1).

Table 1 Position of this review relative to recent reviews of data-centre and immersion cooling. Full coverage is indicated by “yes”, partial or qualitative coverage by “partial”, and absent coverage by “no”.

The gap is then in lack of a common framework rather than lack of information. The use of various coolants, loads, surfaces, scales, and system boundaries makes it hard to establish which architecture is appropriate for a specific heat flux or rack density, what enhancements are hydraulic, and the strength of the evidence regarding long-term operation.

1.4 Novelty, Scope and Contributions

This review addresses these gaps by five connected contributions: the reproducible literature-search and screening process; the five-level chip-to-facility framework; the 12-criterion SPIC/TPIC comparison; harmonised quantitative tables that preserve the test conditions and energy boundaries; and explicit analysis of contradictions, evidence maturity, and deployment readiness.

Immersion cooling of electronic computing equipment for dense and AI systems. Only as a reference architecture, air and direct-to-chip cooling are used. Batteries, photovoltaic modules, and power-electronic applications unrelated to batteries are excluded unless quantitative thermal, hydraulic, energy, economic, or reliability data can be extracted in a relevant manner, as are studies lacking in quantitative thermal, hydraulic, energy, economic, or reliability data.

2. Review Methodology

2.1 Search Strategy

The databases searched were Scopus, Web of Science Core Collection and ScienceDirect, and the last search was made in March 2026. Searches targeted the title, abstract, and keyword fields of all the documents in Table 2, with backward and forward citation chasing. The final corpus consists of 90 studies, out of which 53% were published in 2024-2026.

Table 2 Search protocol: databases, search strings, filters, and inclusion and exclusion criteria applied at each screening stage.

2.2 Inclusion and Exclusion Criteria

We selected studies thatselected if they discussed computing-related immersion cooling (or a directly comparable reference architecture), presented at least one quantitative result, and provided sufficient details about the immersion cooling coolant, heating configuration, and load to be analyzed. Non-computing applications, those without any transferable findings, promotional or descriptive material, incomplete datasets, and superseded conference papers were not included.

2.3 Data Extraction and Harmonisation

Properties of the coolant, heating surface and enhancement, heat load/flux and reference area, flow condition, inlet temperature, h, Rth, ΔT, pressure drop, pumping power, energy metric and boundary, spatial scale, duration, and uncertainty were recorded for each study. Harmonization of heat-flux bases, when possible, and tagging of energy metrics to boundaries B1-B4; no estimation of unreported quantities.

2.4 Quality Appraisal

The completeness of the operating conditions, the reporting of uncertainty, numerical validation, and independence of the performance claims from the promoted enhancement were studied. No study was rejected on quality grounds; appraisal was used to weight the results and identify claims lacking hydraulic or auxiliary-power data.

2.5 Multi-Scale Evaluation Framework

Five scales were used to interpret the results: chip/die, package/heat sink, board/server, tank/rack and facility. Metrics and governing variables vary by scale, and constraints flow both ways: local thermal limits limit the system; and the temperature of heat rejection at a facility and ambient conditions limit chip-level performance. Figure 2 shows the multi-scale evaluation framework used in this review.

Click to view original image

Figure 2 Multi-scale evaluation framework applied in this review.

There are multiple physics associated with each scale, and valid performance metrics for each scale; constraints that move upward from the die and downward from the facility heat-rejection system. Each quantitative finding mentioned in Sections 6-9 is identified as having occurred at which scale.

3. Thermal Requirements of High-Density and AI Computing

With high-density computing, high power and small package sizes challenge density, reliability, and operation [6]. The literature shows five design-relevant differences between existing servers: higher heat flux, higher spatial non-uniformity, multiple interacting heat sources, faster load transients and a much higher power density of the racks (Table 3).

Table 3 Thermal characteristics of modern high-density and AI computing hardware, and their implications for the selection and design of immersion-cooling systems.

3.1 Heat Flux and the Air-Cooling Ceiling

The 700-1400 W dissipations of AI/HPC accelerator packages are approximately 3-10 W cm-2 at the package-average level and local levels even higher. If one chooses 300 W m-2 K-1, then the temperature rise for air becomes too large, but forced SPIC and nucleate boiling can be used in this range of flux. These regimes are plotted on a common heat-flux basis in Figure 3.

Click to view original image

Figure 3 Plotted on the same set of axes, competing cooling mechanisms operate in different regions, (a) the surface heat flux is plotted as a function of the surface-to-fluid temperature difference, with single-phase convective envelopes for the reported ranges of heat-transfer coefficient, nucleate-boiling envelopes for plain and enhanced surfaces bounded by the reported ranges of wall superheat and heat flux, the air-side limit stated in Section 1.1, and the heat-flux bands of mainstream server processors and current AI/HPC accelerators (package average and local hotspots), with forced or jet-assisted SPIC and plain-surface TPIC among the envelopes that intersect these bands; and (b) the heat-transfer coefficient ranges reported from the reviewed studies for each cooling mode. Envelopes are ranges of values derived from the literature reviewed and are not design correlations.

Cooling must provide acceptable temperature control for processors, memory, storage, and power electronics, and to reject heat to the building. So, the maximum temperature, uniformity, the capability of transferring heat, and the cooling energy should be taken into account simultaneously [13].

3.2 Localised Hotspots and Non-Uniform Heat Generation

Temperature control may not be manifested in the package-average h for modern accelerators with local hotspots and dissimilar chiplets. The availability of local coolant is more significant than bulk circulation; component arrangement and inter-board spacing are thermal design variables.

The same applies to adjacent hardware studies: heat spreading can be used to spread the heat of the processor onto larger rejection surfaces, and dense memory arrangements need heat to be spread away from confined areas [14,15]. Both methods are effective in bringing heat to the areas where there is coolant in immersion systems.

Thermal interaction between the accelerators, memory, and voltage-regulation components occurs at the board scale through coolant action. These studies of component arrangement and fin geometry on the server level recently confirmed the importance of server-level evaluation over single-device evaluation [16,17,18,19,20].

3.3 Dynamic Workloads and Transient Behaviour

There is a large power transient associated with AI/HPC workloads, but most immersion testing practices rely on uniform heating. The thermal inertia is large for SPIC, and the response is slow, while the response is near-isothermal for stable boiling and can reach CHF quickly after a load step for TPIC. Future validations should therefore include representative transient workloads.

3.4 Rack-Level Power Density and Facility Coupling

The power of the rack has increased from about 5-15 kW of conventional enterprise systems to 40-130 kW for AI racks. In addition to server-level heat transfer limitations, at this scale, heat rejection, structural loading, fluid inventory, service access and service redundancy can become more limiting.

Therefore, the evaluation criterion for immersion systems should be based on maximum temperature, temperature uniformity, heat-transfer capability, cooling power and equipment density that can be achieved, instead of a single thermal value.

4. Heat-Transfer Mechanisms and Operating Envelopes

First, we set uo governing mechanisms for air and direct-to-chip reference architectures, then for SPIC and TPIC, and finally a common engineering comparison.

4.1 Reference Architectures: Air and Direct-To-Chip Cooling

Air cooling removes heat by circulating air through the server and the room/facility air-handling equipment. Non-uniform airflow can also cause hotspots [7,21], and higher loads require more airflow and fan pressure but consume more infrastructure volume [22].

Direct-to-chip cooling reduces the thermal path of heat through a cold plate and liquid loop, but still has thermal interfaces and typically keeps memory, voltage regulators, and storage on an air path. Performance is thus sensitive to cold-plate design, flow distribution, pressure drop, pump power, and incorporation of the facility [23,24,25,26,27,28,29].

Immersion changes the architecture in a more basic way: immersed components are in direct contact with the dielectric coolant, and are less dependent on server airflow and traditional air-handling systems [30,31,32,33,34].

4.2 Single-Phase Immersion Cooling

In SPIC, the heat is transferred by convection and the coolant temperature stays below its boiling point. A key advantage of natural circulation is that it requires no primary pump, but its velocity is limited by buoyancy and geometry, whereas forced circulation increases local velocity and h but also increases pressure drop and pumping power [30,35].

\[ Q=\dot{m}c_p\Delta T_{coolant} \tag{1} \]

where Q is the heat-removal rate, ṁ is the mass flow rate, cp is specific heat, and ΔT is coolant temperature rise.

\[ Q=hA\Delta T \tag{2} \]

\[ R_{th}=(T_j-T_{coolant})/Q \tag{3} \]

Bulk flow is not enough because of equations (1)-(3) as follows: An increase in ṁ increases transport capacity, but for the local h to be adequate, it must be augmented by flow over high-power surfaces. Thus, the performance of SPIC is influenced by the property of the coolant, geometry of the heat sinks, arrangement and distribution of servers [36,37].

4.3 Two-Phase Immersion Cooling

The boiling of the dielectric in TPIC takes place on heated surfaces; vapor is then transported to the condenser where it condenses and releases the latent heat and flows back to the bath [16,38]. The high heat transfer rates of latent transport depend on the very low temperature differences, but they rely on nucleation, bubble departure, replenishment of liquid, vapour transport, and condensation.

\[ Q=\dot{m}[c_p(T_{sat}-T_{in})+xh_{fg}] \tag{4} \]

For a coolant entering below saturation, Q = ṁ[cp(Tsat - Tin) + xhfg], where Tsat is saturation temperature, x is vapour quality, and hfg is latent heat.

The strong dependence of TPIC on bubble interaction, surface orientation, source spacing, and vapour collection [39,40] is due to multiple heat sources. These effects differ from behaviour in a single-heater boiling situation, especially in terms of server scale.

4.4 Operating Envelopes and the Critical-Heat-Flux Constraint

The operating envelope is the practical difference. SPIC can deliver increasing heat flux, but at increasing pumping penalty and temperature; TPIC can achieve high flux at small superheat but will have a harder CHF limit beyond which rapid temperature rise can result from vapour blanketing.

There is significant overlap in the reported values of h; forced SPIC lies in the range of 0.4-2 kW m-2 K-1, jet/local SPIC 2-15.3 kW m-2 K-1, and plain-surface TPIC 1.5-12 kW m-2 K-1. The next step is to design the boiling surfaces to achieve heat-transfer coefficients of about 5-45 kW m-2 K-1. In fact, the choice of architecture is sensitive not only to CHF margin but also to vapour handling, containment and operating complexity [38,39,41,42,43,44].

4.5 Systematic Comparison of SPIC and TPIC

SPIC conveys sensible heat; TPIC conveys sensible and latent heat by means of boiling and condensation. SPIC is typically easier to operate while TPIC may offer higher flux, but requires stable boiling, vapour management and reliable condensation [16,40].

A comparison of the 12 engineering criteria is continued in Table 4. The main findings are that it is more attractive to be close to the top of the heat-flux envelope with TPIC and that SPIC is more predictable, easier to manage fluid, and has better server/facility evidence. Evidence comparisons should be done on the same scale because there are more SPIC systems at the server scale and more TPIC systems at the chip scale.

Table 4 Comparative analysis of single-phase (SPIC) and two-phase (TPIC) immersion cooling based on 12 engineering criteria, including quantitative ranges, as reported in the literature. Values are indicative ranges determined from the studies conducted in various coolants, surfaces, and scales; the limiting condition is specified when a meaningful numerical range cannot be specified.

5. Dielectric Coolants: An Engineering Selection Perspective

Selecting a dielectric coolant has to take into consideration thermal, hydraulic, electrical, compatibility, safety, economic, and environmental considerations [42,45].

5.1 Coupled Property Requirements

  • In SPIC, conductivity and volumetric heat capacity help to transport heat, and viscosity helps to increase hydraulic demand. Saturation temperature, latent heat, vapour density and surface tension control boiling in TPIC. After ageing, the dielectric properties must be satisfactory, and compatibility with boards, solder, TIMs, polymers, adhesives, seals and elastomers.
  • One of the most important properties that can be used to distinguish SPIC and are significantly influenced by temperature is viscosity. Comparative studies of fluorinated liquids show that differences in thermal-hydraulic properties can follow the viscosity more closely than conductivity [42,46].
  • Coolant choice influences deployability, which is increasingly dependent on fluid inventory, unit cost, global warming potential, PFAS status, biodegradability, and end-of-life route.

5.2 Property Data for the Candidate Fluid Families

In Table 5, representative properties from the literature of the fluid families found in the reviewed literature are compiled, such as mineral and synthetic hydrocarbon oils, natural and synthetic esters, polyalphaolefins, and the hydrofluoroether and perfluorocarbon families of fluids used primarily in TPIC [47]. The values are provided at a specified reference temperature because these properties, especially viscosity, are highly temperature dependent; the table should be used by comparison with families of fluids and the values of the properties of any candidate fluid used should be obtained from the current data sheet of the manufacturer.

Table 5 Some representative characteristics of the families of dielectric coolants for immersion cooling, at about 25°C unless otherwise specified. The values are indicative and are taken from manufacturer data sheets and literature and may differ between product types within the family and with temperature, and should be checked against current data sheets for design purposes. n.a. = not applicable; n.r. = not consistently reported.

Table 5 illustrates the simple trade-off; hydrocarbon/ester fluids generally are more conductive and more heat-capacitive, but also have much higher viscosity, favouring SPIC, whereas the fluorinated fluids have low saturation temperatures favouring TPIC, but lower latent heat and greater fluid-management and environmental constraints.

5.3 Figures of Merit and the Failure of Single-Property Selection

Fluid rank depends on the circulation mode. Forced SPIC gives credit for conductivity and volumetric heat capacity, and penalises the viscosity; buoyancy-driven systems penalise viscosity more heavily. The concepts of latent heat, surface tension and vapour properties come into play in case of boiling, which introduces a different set of merits. Coolant selection needs to be considered a multi-criteria architecture decision and not a ranking of a single property [45,46,48].

5.4 Nanofluids and Modified Coolants

Adding more particles to nanofluids increases effective thermal conductivity; however, viscosity and hydraulic resistance also rise with the number of particles added [49,50].

Other concerns for immersion use include long-term suspension stability, dielectric strength, deposition, loading and pump wear. Thermal and hydraulic drawbacks are moderate, while maintenance drawbacks are significant, keeping nanofluids in the research stage.

5.5 Environmental and Regulatory Constraints

Fluid families also have their environmental and regulatory factors. The GWP and PFAS regulation risk of the volatile fluorinated fluids for TPIC are high, with hydrocarbon and ester fluids posing other ageing and flammability concerns. Fluid loss and end-of-life impacts should, therefore, be part of the life cycle assessment, alongside operational energy.

6. Thermal Performance: Quantitative Synthesis

In an immersion configuration, electronics couple directly to the dielectric coolant, and performance is assessed based on maximum temperature, uniformity, thermal resistance, h, acceptable heat load, and cooling power [41,51]. Sections 6.1-6.4 summarise the quantitative evidence and its paradoxes.

6.1 Single-Phase Immersion Cooling: Reported Thermal-Hydraulic Performance

Local coolant velocity and distribution significantly affect SPIC performance. Low thermal resistance may arise from direct contact, while poor flow paths may result in high power regions under-cooled [35,41,52] (Table 6).

Table 6 Reported performances for single-phase immersion cooling (SPIC) has been quantified, with test conditions noted. The coolant, test article and scale, thermal load, circulation condition and the reported thermal and hydraulic results are stated in each row. If quantities were not reported from the source, they are indicated by the n.r. which is also an observation of this review.

  • Natural-circulation h values are concentrated in the region of 150-500 W m-2 K-1, whereas forced circulation h values are concentrated in the region of 400-2000 W m-2 K-1 and jet/local actuation h values can reach up to 15.3 kW m-2 K-1. The benefits of forced flow must be balanced against pumping power, which is not well documented in the literature.
  • The thermal benefit of a change of the flow through a system, such as providing more equalising plates, baffles or inlet/outlet changes, is greater than increasing the total flow of the system beyond a certain point, as Eq. (2) is a function of local h rather than bulk ṁ.
  • Peak temperature and uniformity are mildly related but very different goals: peak temperature is mostly dependent on local velocity and inlet temperature; uniformity is mostly dependent on flow path, placement of components and depth of immersion.
  • Coolant rankings also change when using different flow modes of circulation, since the importance of viscosity differs between buoyancy and forced flow. The apparent differences between fluid studies may therefore be because they are operating in different regimes, not different physics.

6.2 Two-Phase Immersion Cooling: Reported Boiling Performance

Latent heat provides high local heat transfer rates at almost constant bulk temperature; however, TPIC has a limited range of application beyond CHF and dependent on the coolant, surface, orientation, and operating conditions [39,40,53] (Table 7).

Table 7 A compilation of reported two-phase immersion cooling (TPIC) performance with test conditions labeled: coolant, boiling surface and enhancement, orientation, range of heat flux, superheat of the wall, reported heat transfer coefficient and measured critical heat flux, and system-level metric, where applicable.

  • Surface engineering generally produces larger TPIC gains than switching among fluids within the same family. The h can be raised by a factor of about two to four by porous, microstructured and pin-finned surfaces, as well as by modification of the wettability [31,53,54].
  • The most important TPIC design limit is CHF, but it is less consistently reported than h. AI/HPC heat-flux requirements may overlap with plain-surface CHF, and so it is sometimes necessary to enhance the surface for high-flux TPIC.
  • The differences between multi-chip boiling and single-heater tests are that there is bubble interaction with increased load, vapour accumulation at higher loads, and confined spacing that affects the local heat transfer [39,40].
  • The condenser may become the system limit: the low latent heat of the available dielectric fluids leads to high vapour mass flow, which must be taken into consideration when designing filling ratio, operating pressure, condenser capacity and liquid return with the boiling surface.

6.3 Temperature Uniformity, Hotspot Control and Distributed Components

Immersion depth, heat-source arrangement, flow direction, and coolant distribution are the most important factors for maximum temperature and uniformity at server scale [55,56].

An immersed processor, Memory, voltage regulators, etc. are cooled simultaneously through immersion; however, stagnant areas, close board spacing and inappropriate inlets/outlets can counterbalance direct liquid contact [57,58].

The best conclusion across Table 6 and Table 7 is that thermal performance is related to the spatial relationship between the heat generated and the coolant delivered by the servers. The arrangement of the components, the spacing between boards, obstructions, and inlet/outlet location are, therefore, as important as the properties of the coolant.

6.4 Reconciling Contradictory Findings

The major differences in the literature, shown in Table 8, are methodological rather than physical. Recurring causes are: undeclared energy boundaries, incomplete reporting of properties of the coolant, bulk flow assumed as the local flow of coolant, mismatching of spatial scales, and steady uniform heating applied as realistic workloads.

Table 8 Apparently contradictory findings in the immersion-cooling literature, and their resolution. In both instances, the disagreement can be traced to a lack of a common methodology, not to the physics.

Most apparent contradictions can be attributed to these five causes. Declared boundaries allow energy measures to be compared, complete fluid data allow ranking, local flow measurements remove the need to extrapolate from the chip to the facility, scale tags ensure the metrics are not extrapolated from the chip to the facility, and realistic load profiles reveal transient behaviours.

Overall, thermal resistance decreases, hotspots decrease, and density increases, although this depends on the configuration. Cooling, hydraulics and facility heat rejection should therefore be considered as a whole.

7. Heat-Transfer Enhancement and System Optimization

Enhancement strategies involve increasing local heat transfer, maintaining uniformity and controlling hotspots without causing an undue increase in the pressure drop or auxiliary power requirement. Although Table 9 and Figure 4 compare techniques in terms of net benefit, this is not only in terms of thermal gain.

Click to view original image

Figure 4 (a) Techniques plotted by their reported reduction in peak temperature or thermal resistance against their hydraulic or auxiliary penalty (Table 9), with markers at the mid-points of the reported ranges (bars span the ranges) or, where only an ‘up to’ value is reported, at that upper bound (open markers), gains in boiling heat-transfer coefficient expressed as reductions in boiling resistance, and the size of the marker reflecting the maturity of supporting evidence; techniques to the right of (below) the diagonal are less thermally beneficial than the increase in pumping, auxiliary or actuator power they incur; and the shaded area is for techniques that result in a decrease in auxiliary power. (b) Why single-objective tuning is not good - beyond the knee of the trade-off curve, each additional amount of pumping power will buy minimal temperature margin at the expense of consuming auxiliary energy constantly (schematic illustration with synthetic data; not measured results).

Table 9 Mechanism, thermal benefit, hydraulic or auxiliary energy penalty, scale, and principal limitation and maturity of the supporting evidence for heat-transfer enhancement techniques for immersion cooling.

7.1 Heat-Sink and Fin-Geometry Optimization

The trade-offs in immersion heat sinks are between heat-sink area and flow obstruction. Fin spacing, height, thickness and orientation should be designed to ensure the added area is effectively wetted, as recent oblique and staggered geometries redirect flow, improving temperature uniformity [17,18,19,59,60,61].

7.2 Flow Guiding and Coolant Distribution

To increase the coolant flow to high-power devices and minimize bypass flow, baffles, ducts, channels, and modified inlet arrangements are used [57,58,62]. This tends to yield a higher benefit/penalty ratio compared to increasing total circulation.

7.3 Jet Impingement and Locally Actuated Enhancement

Jet impingement focuses high-momentum coolant on critical surfaces, making it a desirable approach for hig-flux devices [38,43,63,64].

Jet and locally actuated methods can offer some of the highest SPIC local gains (up to h of order 15 kW m-2 K-1) but will need extra power to drive the actuators or pump devices and could introduce moving parts or flow interactions [44,65,66].

7.4 Boiling-Surface Engineering and Wettability Modification

TPIC surface engineering encourages nucleation and keeps the replenishment of the liquid and evacuation of vapours. Increased area and nucleation sites are provided by fins, pin fins, porous structures and microstructures; but their help is dependent on the coupled system of fluid and surface [31,53,54].

Wettability modification is an attractive option as it can improve boiling without a significant hydraulic penalty or auxiliary power, but issues of durability, contamination and reproducibility are not well addressed [53].

7.5 Passive and Hybrid Circulation Strategies

Passive or active circulation trade control against energy use. Natural circulation lowers pump power but generally offers lower capacity, while forced circulation provides higher cooling but higher hydraulic demand [30,67].

If inventory and elevation are matched with the condenser geometry, then the primary pump can be eliminated using a gravity-driven two-phase loop [68,69]. Buoyancy is used at low load, with active flow when needed, to provide reduced auxiliary power and control.

7.6 Multi-Objective Optimization

Immersion optimisation should also try to reduce the maximum temperature, non-uniformity, pressure drop, auxiliary power, and coolant inventory at the same time, not just temperature [70,71].

The trade-off is non-linear: In the beginning, small increases in pumping power or in actuator power can bring substantial thermal improvements, but after a “knee point”, further increments in power will provide minimal improvements in temperature. That is why multi-objective optimisation is better single-objective tuning [72].

8. Energy Efficiency and Facility Integration

Immersion eliminates or reduces the number of server fans, changing the entire cooling chain, but still requires pumps, heat exchangers, condensers (if applicable), and heat rejection from the facility [1,73].

8.1 Metrics, System Boundaries and the Comparability Problem

\[ PUE=P_{total}/P_{IT} \tag{5} \]

where Ptotal is the total facility power and PIT is the power delivered to the IT equipment.

PUE measures facility overhead but does not identify individual cooling contributions.

\[ COP=Q/P_{cooling} \tag{6} \]

where Q is the heat removed by the cooling system and Pcooling is the electrical power consumed by the cooling equipment within the declared boundary.

PUE and COP are only meaningful if there is a declared system boundary. This review is based on B1 server/board, B2 primary loop, B3 tank/rack and B4 facility boundaries (Figure 5).

Click to view original image

Figure 5 Energy metrics and system boundaries: (a) four nested boundaries (B1-B4) used in this review, with the components accounted for in each and the measures (partial PUE or PUE) that are appropriately defined on each; (b) illustrative accounting for one physical system, where the reported measure of PUE or partial PUE increases from 1.036 to 1.142 simply as the accounting boundary expands.

All energy metrics in Table 10 are tagged with the boundary at which they are reported.

Table 10 Results of energy performance reported for the immersion cooling architecture and reference cooling architecture, both labeled with the system boundary B1-B4 on which they were tested, and with the IT load and auxiliary power/heat-rejection configuration that was tested. It is not possible to compare values across different boundaries.

A benefit of immersion is that it eliminates fan energy overhead on the server, and experimental results indicate that immersed power electronics can be more efficient [74].

A study using TPIC reported that as server power was raised from 1127 W to 1577 W, the partial PUE was reduced from 1.053 to 1.037 [73]. These partial-PUE figures are not directly comparable to whole-facility PUE.

If thermal resistance is lower, more flow or more heat-rejection equipment will be needed - and the energy performance needs to take all the additional loads within a declared boundary.

8.2 Heat-Rejection Temperature and Free Cooling

Liquid cooling can reject heat at higher temperatures than chilled-air systems, expanding the potential for economiser and free-cooling.

The usable temperature for heat rejection for immersion depends on the secondary-loop design, ambient temperature and component limits, and the heat-exchanger approach and operating temperature of the coolant [75].

Free-cooling benefit is not a function of climate alone and should be calculated for a representative year, not just for a particular operating point [76].

8.3 Dynamic Operation and Workload Variation

Adaptive operation can adjust coolant flow and heat rejection based on the number of chips being loaded and the ambient conditions, thereby minimizing the use of auxiliary energy at partial load [77].

This is especially relevant for the case of AI/HPC workloads that experience significant fluctuations in the number of processors active and heat generation.

SPIC control is mainly affected by the transport delay & thermal inertia of the coolant, while for TPIC control, the pressure, filling ratio, condenser capacity & CHF margin are important.

8.4 Waste-Heat Recovery

Liquid cooling concentrates waste heat and can facilitate integration with external thermal systems [78,79]. The recovered heat should be considered as useful energy based on the heat-pump power, distribution losses, and temporal mismatch. Higher operating temperatures of the SPIC fluids may allow low-temperature heat reuse without the heat pump, while the saturation temperatures of the common fluorinated fluids limit TPIC.

8.5 Thermal and Energy Storage

Thermal storage and energy storage can help to separate cooling needs from electricity availability. Attempts have been made to use liquid-air cooling storage and other storage concepts to this end [80].

Reservoir thermal energy storage has been related to a summer-peak COP of 16.5 and reductions in electricity consumption, cost and emissions for a modelled 5 MW cooling load [81].

Energy storage round-trip efficiency for a waste heat-assisted transcritical CO2 system is reported to be approximately 64.67-67.41% [82].

Storage strategies can serve several purposes, including storing cooling capacity and/or recovering rejected heat, and can be valuable depending on capital cost, losses, and system complexity.

9. Economic and Environmental Assessment

Deployment capability is not solely thermal. Immersion equipment, avoided air infrastructure, coolant inventory, maintainability, replacement, containment and end-of-life costs are all necessary parts of the economics.

9.1 Cost Structure

Avoidance of fans, air handlers, containment and chiller infrastructure (where appropriate) in new-built facilities can be significant, and the economic case is strongest in these situations. The fluid stock inventory is a capital cost that increases with tank volume, and there is limited quantification of replacement, maintenance or disposal costs. The principal cost items, their direction relative to an air-cooled baseline and their coverage in the literature are summarised in Table 11.

Table 11 The differences in the immersion cooling and air-cooled baseline cost items and their direction, their main drivers, and their literature coverage.

9.2 Reported Savings and Payback

Cooling-electricity savings need to be balanced against new liquid-cooling equipment costs. The result depends heavily on climate, IT load, coolant temperature, heat rejection, auxiliary equipment, and economisation.

Liquid cooling studies have been reported, and potentially significant energy savings have been achieved, depending on the baseline and climate. The other connected cooling study indicated a 2.48-6.64 years dynamic payback, and 34.9-52.4% annual energy savings [83,84,85].

Payback values are indicative only, as tariffs, climate and service-life costs for excluded services, particularly fluid replacement and disposal, will significantly influence them.

9.3 Fluid Inventory, Replacement and Containment Cost

Significant fluid inventories are required for large immersion systems, influencing tank size, structural loading, initial system costs and servicing.

With volatile fluid coolants, TPIC is more likely to experience recurrent fluid loss exposure due to the volatilization of fluid (loss during maintenance and/or through containment pathways), and in the case of high fluid cost, intervention frequency becomes an economic decision.

9.4 Environmental Impact and End of Life

During the design of operations, consider energy, water, fluid GWP/PFAS status, make-up fluid treatment, and end-of-life treatment. Comparisons of operational energy alone may bias the results in favour of TPIC, which do not include the burden of fluid [86].

Regulatory uncertainty for PFAS is independent of thermal performance and strengthens the case for compliant TPIC fluids.

10. Long-Term Reliability and Deployment Readiness

Thermal capability is well proven at chip-to-server scale, but only multi-year time series of data exsit on fluid and hardware survival. Short laboratory tests do not capture the slow failure modes of materials, fluids and seals.

10.1 Material Compatibility

Immersion means continuous contact with the coolant of PCBs, connectors, solder, TIMs, cable insulation, polymers, adhesives, seals and elastomers. Failure modes include swelling or extraction of plasticiser, TIM degradation, solder/intermetallic attack, coating removal and corrosion, which can affect mechanical, electrical, or thermal properties over months to years [46].

10.2 Coolant Ageing and Contamination

Thermal cycling, oxidation, moisture absorption, contamination, dissolved material and interaction with hardware [42,45,46] can change coolant properties.

Dielectric strength, moisture, acid number, viscosity, particles and dissolved metals are the major ageing indicators. Oxidation can increase viscosity and acidity, moisture can lead to a lowering of the dielectric strength, deposits can contaminate surfaces, and local thermal decomposition can lead to corrosive products.

10.3 Available Long-Duration Evidence and Its Limits

Long-duration testing should not only validate the thermal performance, but also provide evidence of the stability of the fluid and hardware under realistic service conditions.

The five-month mineral oil and virgin coconut oil exposure study showed no damage to the components, but it is not generalizable due to the duration and fluid set used [87] (Table 12).

Table 12 Evidence of immersion cooling from the reviewed corpus: test duration, type of fluid, hardware that was immersed, quantities that were monitored, reported outcome, limitation that does not allow for generalization.

The longest exposure found was five months, whereas the hardware service lives are multi-year. Few reports describe the evolution of dielectric strength, acid number, elastomer properties and in-service fluid condition for multi-year operation.

10.4 Hydraulic, Containment and Scale-Up Constraints

The distribution of flow is influenced by forced SPIC by the presence or absence of adequate coolant delivery, inlet/outlet arrangement, board orientation, spacing, heat sinks and obstructions which produce the flow field [48,88].

If distribution is poor, the capacity of the pump can be increased to a certain degree, but it can also increase auxiliary energy consumption; in general, geometric flow management is better.

Stable nucleation, vapour removal, condensation and liquid return must be maintained in TPIC. The effect of the interaction between bubbles and the accumulation of vapour increases as the load increases [39,40,89,90].

The condenser capacity also depends on the vapour generator capacity and cannot be determined solely by the size of the boiling surface.

Scale-up coolant distribution, tank volume, structural loading, heat exchanger, secondary loops and redundancy and service access [33,36,88].

The extent of evidence at multi-server and facility scale is considerably less advanced than evidence at the component level.

Serviceability is also undervalued: draining, drip-off, cleaning, lifting and, in the case of TPIC, fluid loss affect maintenance and total ownership cost.

The key deployment issues for SPIC are flow distribution and pumping energy, whereas stable boiling, vapour handling, condensation and containment are the key issues for TPIC. Both need to be reliable in heat rejection, monitoring, maintenance and redundancy.

11. Research Gaps, Agenda and Reporting Requirements

11.1 Where the Evidence Is and Is Not Sufficient

There is a highly imbalanced evidence base for each topic and spatial scale. Thermal performance and enhancement are relatively well developed at the chip, package, and server scale, and partial to well developed at a small scale but sparse at a tank scale and absent at a facility scale, with regard to coolant thermophysics. Energy efficiency, facility coupling and techno-economics are partial at tank and facility scale, and almost non-existent at component scale due to the lack of energy reporting in studies at this scale. Only a few of the scales reviewed include indications for material compatibility, multi-year coolant ageing, or standardized test protocols. The bottom line is that the field has developed a good quantity of evidence on the effectiveness of immersion cooling for heat removal, but much less evidence on the performance of an immersion installation over several years. The latter would entail long-duration exposure data, property evolution of the coolant and materials, and standardisation of test protocols. The gaps are summarized in Table 12 and converted to specific research questions and needed approaches in Table 13.

Table 13 Research gaps identified in this review, sorted by level of observation, the question to be answered, the experimental or analytical method needed, and priority assigned.

11.2 Prioritised Research Agenda

Near-term priorities are simple but consequential: report IT load, Auxiliary power, and System boundary for each energy metric; report full coolant properties at operating temperature; report pressure drop or actuator power for each enhancement result.

Medium-term priorities are a round-robin immersion test protocol, matched server-to-rack scaling studies and representative AI/HPC transient load profiles.

Long-term priorities are multi-year fluid-hardware ageing campaigns, PFAS-compliant boiling fluids and facility demonstrators with audited energy and water and cost accounting.

11.3 Minimum Reporting Requirements for Future Studies

The minimum reporting standard should include: coolant and properties, surface preparation, heat load and flux basis, flow conditions, temperature measurements, pressure drop and auxiliary powers, spatial scale, source count, duration and uncertainty. In addition, energy studies should include IT energy loads, all auxiliary energy loads, energy rejected, and the complete system boundary.

12. Conclusions

The 90 studies included in this review were selected from the published literature and span 2015 to 2026, following a documented search protocol, a five-level scale framework, a 12-criterion SPIC/TPIC comparison and harmonised quantitative tables. The main conclusions are dependent on the scale.

The air-side limitation is eliminated at the chip/package scale. Forced SPIC (≈0.4-2 kW m-2 K-1) and plain-surface TPIC (≈1.5-12 kW m-2 K-1) cover the heat-flux range of current AI/HPC accelerators, while jet SPIC and engineered boiling surfaces reach the high limits. Thus, the choice of architecture is dictated by CHF margin, vapour handling, containment, and operating complexity as well as h.

At the server/board level, the most important variable is the heat-to-coolant balance. Beyond increasing total flow, these factors can achieve better thermal uniformity more efficiently: flow redistribution, inlet/outlet design, fin arrangement, and component arrangement.

In the case of tank/facility scale, the boundaries of PUE and COP need to be clear. Energy, heat rejection, climate, cost, water, and IT load need to be evaluated as a single system because the reported values for these quantities will vary depending on the accounting boundary.

Overall, SPIC is the more predictable and easier path to deployment, and has fully developed hydraulic design, as well as a wider base of server/facility evidence. TPIC provides higher high-flux, near-isothermal boiling, and introduces CHF, vapour, condenser, containment, fluid-cost/regulatory constraints.

The biggest remaining gap is deployment evidence - multi-year ageing, material compatibility, standardised testing, operation at the facility level, with audited results. What’s needed right now is not more one-off, peak performance showcases, but disciplined, similar reporting and long-duration validation.

Nomenclature

Abbreviations

Author Contributions

Ghassan Adnan Abd: conceptualization of the review framework, literature search and screening, data extraction and harmonisation, preparation of the figures and tables, and writing of the original draft. Ahmed J. Shkarah: methodology of the search protocol and quality appraisal, verification of the extracted quantitative data, and review and editing of the manuscript. Mustafa M. Mansour: conceptualization, supervision and project administration, critical analysis and interpretation of the comparative findings, and review and editing of the final manuscript. All authors have read and approved the final version of the manuscript.

Image Processing Statement

All figures in this article (Figures 1-5) are original graphics prepared by the authors; no photographs, micrographs or third-party images were used, reproduced or adapted. All five figures were produced programmatically in Python using the Matplotlib library: the schematic diagrams (Figure 1, Figure 2 and Figure 5(a)) were built from Matplotlib shape patches, arrows and text annotations, and the plots (Figure 3, Figure 4 and Figure 5(b)) were generated with Matplotlib line, bar, area-fill and scatter functions; Figure 4(b) and Figure 5(b) are illustrative examples, as stated in the figures. Figures 3 and 4(a) are plotted directly from data files containing the values given in Tables 3, 4, 7 and 9 and the air-side limit stated in Section 1.1; quantities derived from these values (the hotspot band, mid-points of reported ranges, upper bounds and the conversion of boiling heat-transfer gains) follow rules stated in the figures and the data files. The box widths in Figure 1 are qualitative, indicative shares of the junction-to-ambient resistance rather than computed values. The layouts of all five figures, and the indicative or illustrative values in Figures 1, 4(b) and 5(b), are based on earlier versions of the figures, which had been generated with Claude (Anthropic) from the authors’ descriptions and have now been replaced. The Python scripts and data files, which regenerate all five figures, were prepared with the assistance of Claude and are available from the corresponding author. Each figure was exported from Matplotlib as a PNG image at 600 dpi and at the text width, and was inserted into the manuscript at that size without cropping. No part of any image was modified after export: no element was retouched, cloned, erased, added or moved, and no brightness, contrast or colour adjustment was applied to the whole or any part of any figure.

Competing Interests

The authors have declared that no competing interests exist.

AI-Assisted Technologies Statement

Artificial intelligence (AI) tools were used in the preparation of this manuscript for two purposes. OpenAI’s ChatGPT was employed for basic grammar correction and language refinement, to improve the readability and linguistic clarity of the English text. Claude (Anthropic) was used to assist in preparing the Python scripts and data files from which all five figures are produced and in drafting the revised figure captions and figure-related statements; earlier versions of the figures that had been generated with Claude have been replaced (see the Image Processing Statement). All scientific content, data interpretation, and conclusions were developed independently by the authors. The authors have thoroughly reviewed and edited the AI-assisted text to ensure its accuracy and accept full responsibility for the content of the manuscript.

References

  1. Zhang C, Sun X, Han Z, Li X, Dong J. Energy saving potential analysis of two-phase immersion cooling system with multi-mode condenser. Appl Therm Eng. 2023; 219: 119614. [CrossRef] [Google scholar]
  2. Shrigondekar H, Lin YC, Wang CC. Investigations on performance of single-phase immersion cooling system. Int J Heat Mass Transfer. 2023; 206: 123961. [CrossRef] [Google scholar]
  3. Taddeo P, Romaní J, Summers J, Gustafsson J, Martorell I, Salom J. Experimental and numerical analysis of the thermal behaviour of a single-phase immersion-cooled data centre. Appl Therm Eng. 2023; 234: 121260. [CrossRef] [Google scholar]
  4. Cai S, Gou Z. Towards energy-efficient data centers: A comprehensive review of passive and active cooling strategies. Energy Built Environ. 2024; 7: 206-226. [CrossRef] [Google scholar]
  5. He W, Li C, Xu C, Yang Y, Liu S, Cao B, et al. Performance of air-liquid integrated cold plate heat sinks in servers. Appl Therm Eng. 2025; 281: 128734. [CrossRef] [Google scholar]
  6. Al Kez D, Foley AM, Wong FW, Dolfi A, Srinivasan G. AI-driven cooling technologies for high-performance data centres: State-of-the-art review and future directions. Sustain Energy Technol Assess. 2025; 82: 104511. [CrossRef] [Google scholar]
  7. Omrani M, Ghassemi M. AI-driven optimization of fan-wall cooling system in a medium-density data center. Int J Heat Mass Transfer. 2025; 247: 127159. [CrossRef] [Google scholar]
  8. Yang Z, Wang Y, Yuan X, Wang X, Yang X, Li Q. Design and optimization of a staggered counter-current microchannel liquid-cooled plate. Appl Therm Eng. 2025; 279: 127891. [CrossRef] [Google scholar]
  9. Sun X, Han Z, Li X. Simulation study on cooling effect of two-phase liquid-immersion cabinet in data center. Appl Therm Eng. 2022; 207: 118142. [CrossRef] [Google scholar]
  10. Zhou K, Yu X, Xie B, Xie H, Fu W. Immersion cooling technology development status of data center. Sci Technol Energy Transit. 2024; 79: 41. [CrossRef] [Google scholar]
  11. Zheng S, Su C, Yang X, Zhang Y, Duan K, Zhang Y, et al. A comprehensive review of single-phase immersion cooling in data centres. Appl Therm Eng. 2025; 272: 126385. [CrossRef] [Google scholar]
  12. Lua KB, Wang CC. Single-phase immersion cooling for data centers and HPC: Recent advances and system-level challenges. Renew Sustain Energy Rev. 2026; 232: 116786. [CrossRef] [Google scholar]
  13. Xue ZH, Ai BC, Qu W. Data center energy conservation study utilizing loop heat pipes as a chip-level cooling technique and its industrial application. Appl Therm Eng. 2023; 220: 119715. [CrossRef] [Google scholar]
  14. Shah D, Pandya N, Kim SK, Mistofsky K, Eamons A. Thermal impact of ultra-narrow DIMM pitch and air-cooled heat pipe solution on optimizing DDR5 cooling in high density HPC systems. Therm Sci Eng Prog. 2025; 68: 104206. [CrossRef] [Google scholar]
  15. Li C, Li J. Thermal characteristics of a flat plate pulsating heat pipe module for onsite cooling of high power server CPUs. Therm Sci Eng Prog. 2023; 37: 101542. [CrossRef] [Google scholar]
  16. Khoshvaght-Aliabadi M, Ghodrati P, Shin JY, Kang YT. Impact of coolant distribution design on server-level thermal management in data centers. Energy. 2025; 330: 136936. [CrossRef] [Google scholar]
  17. Khoshvaght-Aliabadi M, Ghodrati P, Nasrolahzadeh A, Kang YT. Optimization of heat sinks for data center server CPUs cooled via single-phase immersion cooling. Appl Therm Eng. 2025; 280: 127790. [CrossRef] [Google scholar]
  18. Cheng SW, Zhang YD, Chien LH, Wang CC. A comparative study of a single-phase immersion-cooled server with a pin-fin heat sink for mitigation of the flow bypass effect. Processes. 2026; 14: 2209. [CrossRef] [Google scholar]
  19. Khoshvaght-Aliabadi M, Nasrolahzadeh A, Ghodrati P, Kang YT. Oblique-fin heat sink design for uniform CPU cooling in immersion-cooled servers. Appl Therm Eng. 2025; 288: 129628. [CrossRef] [Google scholar]
  20. Khoshvaght-Aliabadi M, Hojjati F, Hassani S. Influence of component arrangement on thermal management in immersion-cooled server boards. Energy. 2026; 347: 140323. [CrossRef] [Google scholar]
  21. Liu P, Kandasamy R, Ho JY, Wong TN, Toh KC. Dynamic performance analysis and thermal modelling of a novel two-phase spray cooled rack system for data center cooling. Energy. 2023; 269: 126835. [CrossRef] [Google scholar]
  22. Cho J. Numerical analysis of rack-based data center cooling with rear door heat exchanger (RDHx): Interrelationship between thermal performance and energy efficiency. Case Stud Therm Eng. 2024; 63: 105247. [CrossRef] [Google scholar]
  23. Heydari A, Gharaibeh AR, Tradat M, Manaserh Y, Radmard V, Eslami B, et al. Experimental evaluation of direct-to-chip cold plate liquid cooling for high-heat-density data centers. Appl Therm Eng. 2024; 239: 122122. [CrossRef] [Google scholar]
  24. Hu H, Chen C, Li C, Pan M. Experimental investigation of roll bond liquid cooling plates for server chip heat dissipation. Appl Therm Eng. 2023; 226: 120284. [CrossRef] [Google scholar]
  25. Samal SK, Chang HC, Fulpagare Y, Wang CC. Thermal management of data centers: Chip-scale cooling using novel distributed inlet-outlet jet impingement liquid cold plate. Appl Therm Eng. 2025; 271: 126360. [CrossRef] [Google scholar]
  26. Jiao J, Duan F, Gong L. Multi-objective optimal design of hybrid liquid cooling system with manifold microchannels for data center server thermal management. Energy. 2025; 337: 138648. [CrossRef] [Google scholar]
  27. Cho J, Moon JH. Numerical coupling of energy efficiency and thermal performance for cold plate cooling optimization in high-density compute AI data centers. Energy Build. 2025; 348: 116441. [CrossRef] [Google scholar]
  28. Li J, Luo X, Wang M, Chen C, Yang C, Zhang Z, et al. Study of flow and heat transfer characteristics of tandem cold plates for data center cooling. Case Stud Therm Eng. 2025; 73: 106590. [CrossRef] [Google scholar]
  29. Qu S, Wang C, Liu Y, Feng Y, Han Y, Xing Z. System design, control strategy development and performance comparison analysis for a novel air-liquid hybrid data center cooling system based on cold plates. Int J Refrig. 2025; 180: 222-239. [CrossRef] [Google scholar]
  30. Huang Y, Ge J, Chen Y, Zhang C. Natural and forced convection heat transfer characteristics of single-phase immersion cooling systems for data centers. Int J Heat Mass Transf. 2023; 207: 124023. [CrossRef] [Google scholar]
  31. Shete RK, Laskar N, Das MK. Geometry-driven high-density heat sink module for enhanced two-phase immersion cooling. Therm Sci Eng Prog. 2025; 68: 104237. [CrossRef] [Google scholar]
  32. Liu Z, Sun X, Ji S, Yuan K. A multi-objective optimization of a single-phase immersion cooling system at cabinet level. Energy Build. 2024; 312: 114221. [CrossRef] [Google scholar]
  33. Kim J, Choi H, Lee S, Lee H. Computational study of single-phase immersion cooling for high-energy density server rack for data centers. Appl Therm Eng. 2025; 264: 125476. [CrossRef] [Google scholar]
  34. Wang J, Yuan C, Li Y, Li C, Wang Y, Wei X. Direct-to-chip immersion liquid cooling for high-power vertical-cavity surface-emitting laser (VCSEL). Appl Therm Eng. 2025; 269: 126137. [CrossRef] [Google scholar]
  35. Sun X, Liu Z, Ji S, Yuan K. Experimental study on thermal performance of a single-phase immersion cooling unit for high-density computing power data center. Int J Heat Fluid Flow. 2025; 112: 109735. [CrossRef] [Google scholar]
  36. Lionello M, Rampazzo M, Beghi A, Varagnolo D, Vesterlund M. Graph-based modelling and simulation of liquid immersion cooling systems. Energy. 2020; 207: 118238. [CrossRef] [Google scholar]
  37. Gheni M, Kerskes H, Stergiaropoulos K. 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. Appl Energy. 2026; 403: 127061. [CrossRef] [Google scholar]
  38. Liu C, Huang Y, Zhang C. Efficient jet-assisted single-phase immersion liquid cooling for high heat-flux servers. Appl Therm Eng. 2025; 259: 124935. [CrossRef] [Google scholar]
  39. Krishnadasan VB, Suresh P, Balaji C. Experimental investigations on the performance and bubble dynamics of two-phase immersion cooling system with multiple chips. Int J Heat Mass Transf. 2025; 245: 126977. [CrossRef] [Google scholar]
  40. Wu X, Yang J, Liu Y, Zhuang Y, Luo S, Yan Y, et al. Investigations on heat dissipation performance and overall characteristics of two-phase liquid immersion cooling systems for data center. Int J Heat Mass Transf. 2025; 239: 126575. [CrossRef] [Google scholar]
  41. Huang Y, Liu B, Xu S, Bao C, Zhong Y, Zhang C. Experimental study on the immersion liquid cooling performance of high-power data center servers. Energy. 2024; 297: 131195. [CrossRef] [Google scholar]
  42. Chen X, Huang Y, Xu S, Bao C, Zhong Y, Chen Y, et al. Thermal performance evaluation of electronic fluorinated liquids for single-phase immersion liquid cooling. Int J Heat Mass Transf. 2024; 220: 124951. [CrossRef] [Google scholar]
  43. Huang Y, Liu C, Zhong Y, Zhang C. Experimental study on jet-enhanced immersion liquid cooling for energy-efficient data centers. Energy. 2025; 334: 137584. [CrossRef] [Google scholar]
  44. Azarifar M, Arik M. Liquid synthetic jets for high flux electronics cooling. Appl Therm Eng. 2025; 261: 125007. [CrossRef] [Google scholar]
  45. Li X, Guo S, Sun H, Liu S, Wang X, Wang L. Assessment of coolants used single-phase immersion cooling data center. Energy. 2025; 340: 139240. [CrossRef] [Google scholar]
  46. Ruichek A, Hnayno M, Chehade A, Fardoun F, Moussa T, Polidori G, et al. Immersion cooling fluid comparison using a single-phase immersion/liquid technique for data centers. Appl Therm Eng. 2025; 280: 128355. [CrossRef] [Google scholar]
  47. Koucheh AB, Sharbati P, Ünlü C, Koşar A, Sadaghiani A. Comprehensive evaluation of battery cooling mechanisms including two-phase immersion with 3M™ NOVEC™-7000 and 7100. Energy Convers Manage. 2026; 348: 120780. [CrossRef] [Google scholar]
  48. Rongyang Y, Yu Z, Liang R, Su W, Wei J. Numerical investigation on thermal performance of single-phase immersion cooling systems using oil coolant. Front Heat Mass Transfer. 2025; 23: 279-298. [CrossRef] [Google scholar]
  49. Luo Q, Wang C, Wen H, Liu L. Research and optimization of thermophysical properties of sic oil-based nanofluids for data center immersion cooling. Int Commun Heat Mass Transf. 2022; 131: 105863. [CrossRef] [Google scholar]
  50. Alkasmoul FS, Almogbel AM, Shahzad MW, Al-Damook AJ. Thermal performance and optimum concentration of different nanofluids in immersion cooling in data center servers. Results Eng. 2025; 25: 103699. [CrossRef] [Google scholar]
  51. Gonzalez RC, Diaz AJ, Rojas D, Sarmiento-Laurel C. Assessing solar absorption chillers for data center cooling across diverse climatic conditions. Energy Build. 2025; 331: 115376. [CrossRef] [Google scholar]
  52. Muneeshwaran M, Lin YC, Wang CC. Performance analysis of single-phase immersion cooling system of data center using FC-40 dielectric fluid. Int Commun Heat Mass Transf. 2023; 145: 106843. [CrossRef] [Google scholar]
  53. Kalantarpour R, Vafai K. Investigating the two-phase immersion cooling performance by surface wettability modification. Int J Heat Mass Transf. 2026; 255: 127796. [CrossRef] [Google scholar]
  54. Zhou G, Zhou J, Huai X, Zhou F, Jiang Y. A two-phase liquid immersion cooling strategy utilizing vapor chamber heat spreader for data center servers. Appl Therm Eng. 2022; 210: 118289. [CrossRef] [Google scholar]
  55. Ding B, Xia F, He L, Huang Y. A novel concept for performance enhancement of immersion-cooled data center servers. Int Commun Heat Mass Transf. 2024; 159: 107980. [CrossRef] [Google scholar]
  56. Suresh P, Chang KW, Lua KB, Wang CC. Deep reinforcement learning for energy-efficient thermal management in 2U air-cooled server systems. Appl Energy. 2026; 404: 127168. [CrossRef] [Google scholar]
  57. Ni D, Yu F. CFD simulation study of flow equalisation plate model in single-phase immersion liquid cooling for servers. Therm Sci Eng Prog. 2024; 47: 102268. [CrossRef] [Google scholar]
  58. Wang H, Yuan X, Zhang K, Lang X, Chen H, Yu H, et al. Performance evaluation and optimization of data center servers using single-phase immersion cooling. Int J Heat Mass Transf. 2024; 221: 125057. [CrossRef] [Google scholar]
  59. Ozer RA. Heat sink optimization with response surface methodology for single phase immersion cooling. Int J Heat Fluid Flow. 2025; 112: 109745. [CrossRef] [Google scholar]
  60. Li X, Xu Z, Liu S, Zhang X, Sun H. Server performance optimization for single-phase immersion cooling data center. Appl Therm Eng. 2023; 224: 120080. [CrossRef] [Google scholar]
  61. Pandey V, Liang SS, Law LL, Kuppusamy NR, Ching-Chiuan Y, Lee PS, et al. Experimental evaluation of 3D printed topology optimized liquid and hybrid-cooled heat sinks for data centers. Int J Heat Mass Transf. 2026; 256: 128106. [CrossRef] [Google scholar]
  62. Xu Y, Yang Y, Tian M, Li D, Deng T, Hu K, et al. An investigation into the effects of various IT scheduling algorithms on thermal control and energy efficiency in liquid-cooled data centers. J Build Eng. 2025; 107: 112798. [CrossRef] [Google scholar]
  63. Huang Y, Liu C, Ding B, Deng Z, Gao W. Machine learning-optimized jet-enhanced immersion liquid cooling for high-power data centers. Int Commun Heat Mass Transf. 2025; 169: 109654. [CrossRef] [Google scholar]
  64. Jin P, Zhang X, Zhang X, Tan Z, Xie G. Immersed jet impingement heat transfer for servers in immersion cooling data center: A parametric study. Appl Therm Eng. 2025; 274: 126628. [CrossRef] [Google scholar]
  65. He W, Liu C, Du L, Yang Y, Liu S, Zeng M, et al. Structural optimization on vertical jet impingement cooler used for sever heat dissipation in data centers. Appl Therm Eng. 2025; 280: 128432. [CrossRef] [Google scholar]
  66. Jin P, Yang J, Zhang X, Zhang X, Xie G. Heat transfer enhancement for high power chips in immersion cooling by ducted propellers. Appl Therm Eng. 2025; 273: 126469. [CrossRef] [Google scholar]
  67. Charaf H, Maalouly M, Chehade A, Al Moussawi H, Moussa T, Polidori G, et al. Passive hydraulic cooling: A CDU-free solution for data center efficiency. Appl Therm Eng. 2025; 280: 128116. [CrossRef] [Google scholar]
  68. Wang L, Chen L, Wang X, Zheng H, Chen X, Li Z. Design and experimental study of a gravity-driven two-phase loop system for multi-elevation heat source cooling. J Energy Storage. 2026; 142: 119517. [CrossRef] [Google scholar]
  69. Zhang Y, Gao Y, Wang Y, Ma X, Yang X, Wei J. Immersion phase-change thermal management system coupled with loop thermosyphons for utilization in data centers cooling. Energy Convers Manage. 2026; 348: 120688. [CrossRef] [Google scholar]
  70. Zou S, Zhang J, Wang Y, Xiao T, Huang J. Multi-objective optimization of server component temperatures and coolant filling ratio in single-phase immersion tank based on Taguchi grey correlation analysis. Case Stud Therm Eng. 2025; 72: 106363. [CrossRef] [Google scholar]
  71. Liu HM, Chang Z. Multi-objective optimization of temperature uniformity in the immersion liquid cooling cabinet with Taguchi-based grey relational analysis. Int Commun Heat Mass Transf. 2024; 154: 107395. [CrossRef] [Google scholar]
  72. Qin X, Shen A, Duan H, Liu H, Wang D, Liu H, et al. Investigating 4E analyses and multi-objective optimization of a novel data center compression/ejection transcritical CO2 cooling system. Appl Therm Eng. 2025; 264: 125477. [CrossRef] [Google scholar]
  73. Ma Y, Bao Y, Li J. Heat transfer dependence of power usage effectiveness of an augmented two-phase immersion cooling system for high-power servers. Energy. 2025; 323: 135853. [CrossRef] [Google scholar]
  74. Liu X, Li M, Wang F, Zhang H, Wang H, Wen B, et al. Thermal performance analysis of single-phase immersion cooling for power device: An experimental study. Appl Therm Eng. 2025; 284: 129142. [CrossRef] [Google scholar]
  75. He W, Zhang J, Li H, Liu S, Wang Y, Lv B, et al. Optimal thermal management of server cooling system based cooling tower under different ambient temperatures. Appl Therm Eng. 2022; 207: 118176. [CrossRef] [Google scholar]
  76. Liu S, Guo S, Sun H, Xu Z, Li X, Wang X. Evaluation of energy, economic, and pollution emission for single-phase immersion cooling data center with different economizers. Appl Therm Eng. 2025; 260: 125049. [CrossRef] [Google scholar]
  77. Wang C, Yu H, Xian L, Liu G, Chen L, Tao WQ. Study on the operation strategy of the water cooling system in the single-phase immersion cooling data center considering cooling tower performance. Energy Build. 2024; 324: 114890. [CrossRef] [Google scholar]
  78. Huang G, Xie M, Zhao Y, Jia T, Dai Y. Experimental study on waste heat recovery process in a water-cooled data center for dehumidification application using desiccant-coated heat exchangers. Energy Build. 2025; 346: 116196. [CrossRef] [Google scholar]
  79. Fu C, Zhang W, Zhou X, Shen Q, Wu T. Stochastic optimization of photovoltaic-integrated data centers with hybrid cooling and waste heat recovery for district energy supply. Renew Energy. 2025; 256: 124248. [CrossRef] [Google scholar]
  80. Liu C, Hao N, Zhang T, Wang D, Li Z, Bian W. Optimization of data-center immersion cooling using liquid air energy storage. J Energy Storage. 2024; 90: 111806. [CrossRef] [Google scholar]
  81. Oh H, Jin W, Peng P, Winick JA, Sickinger D, Sartor D, et al. Techno-economic performance of reservoir thermal energy storage for data center cooling system. Appl Energy. 2025; 391: 125858. [CrossRef] [Google scholar]
  82. Li F, Li P, Ding R, Su W, Ou S, Xing L, et al. Thermo-economic analysis on trans-critical compressed CO2 energy storage system integrated with the waste heat of liquid-cooled data center. J Energy Storage. 2024; 103: 114292. [CrossRef] [Google scholar]
  83. Ren H, Li Y, Lin W, Sun Y. Energy and economic analysis of photovoltaic thermal night cooling and adaptive reset control for ultra-efficient data center immersion cooling. Renew Energy. 2025; 255: 123821. [CrossRef] [Google scholar]
  84. Du S, Zhang Q, Zou S, Meng F, Liu L. Simulation analysis on energy consumption and economy of CPU cooling system based on loop heat pipe for data center. Therm Sci Eng Prog. 2023; 45: 102115. [CrossRef] [Google scholar]
  85. Kanbur BB, Wu C, Fan S, Tong W, Duan F. Two-phase liquid-immersion data center cooling system: Experimental performance and thermoeconomic analysis. Int J Refrig. 2020; 118: 290-301. [CrossRef] [Google scholar]
  86. Zhang M, Carbajales-Dale M, Ma X, Guo L, Fan C. Cleaner grid or smarter cooling? Environmental impact trade-offs of a data center using the life cycle assessment method. Clean Energy Syst. 2025; 12: 100223. [CrossRef] [Google scholar]
  87. Pambudi NA, Yusuf AM, Sarifudin A. The use of single-phase immersion cooling by using two types of dielectric fluid for data center energy savings. Energy Eng. 2021; 119: 275-286. [CrossRef] [Google scholar]
  88. Liu S, Xu Z, Wang Z, Li X, Sun H, Zhang X, et al. Optimization and comprehensive evaluation of liquid cooling tank for single-phase immersion cooling data center. Appl Therm Eng. 2024; 245: 122864. [CrossRef] [Google scholar]
  89. He Z, Smith TW, Lei Z, Taleghani AD. Repurposing abandoned mine lands for cooling data centers. Resour Conserv Recycl. 2026; 225: 108640. [CrossRef] [Google scholar]
  90. Ejaz F, Kwon B. Two-phase active immersion cooling for vertically mounted electronics with interchip component-assisted bubble departure. Int Commun Heat Mass Transfer. 2024; 159: 107981. [CrossRef] [Google scholar]
Newsletter
Download PDF Download Citation
0 0
Newsletter  

TOP