CFD simulation for data centers

Almost all the electricity entering a data center leaves it as heat. The subject is therefore not "producing cold", but removing a thermal power that is increasingly concentrated in a volume which, for its part, has not grown.

Demand for digital infrastructure has changed order of magnitude. According to the International Energy Agency, data centers consumed around 415 TWh in 2024, close to 1.5 % of global electricity consumption, growing by about 12 % per year since 2017; the base case takes them to 945 TWh in 2030, a doubling in six years. That growth does not only mean more buildings: it means more watts per square metre of floor.

This is where the engineering break lies. A room designed around 5 kW per rack tolerates many approximations: the air volume of the room acts as a buffer, distribution errors are diluted, and a cooling failure leaves several tens of minutes of margin. At 15, 30 or 60 kW per rack, the same 10 % distribution error no longer produces one degree of deviation but five, the slightest hot air leak becomes a hot spot, and the time available before manufacturer thresholds are exceeded is counted in minutes, sometimes tens of seconds.

Field surveys confirm that the installed base is moving more slowly than the announcements: the Uptime Institute observes a continuous but slow rise in average densities, with growing adoption of the 10–30 kW range, one site in eight reporting racks between 30 and 59 kW, and a few cabinets beyond 100 kW still exceptional. This heterogeneity is precisely the problem: most operators today have to insert high-density islands into rooms designed for low, homogeneous density. It is the most frequent use case for CFD.

Concrete consequences of poorly controlled thermals

• Permanent operating overcost Preventive overcooling is paid for every hour of the year, in cooling production and fan energy.
• Accelerated ageing Thermal cycles and high intake temperatures degrade electronics, fans and batteries.
• Shutdown risk Beyond the thresholds, servers reduce their frequency (throttling) then stop: often several racks at once, because they share the same airflow defect.
• Frozen capacity For want of thermal proof, physically available positions remain off limits to load. That is a direct loss of revenue.

Classic methods, such as the global power balance, the "total airflow is sufficient" rule or the per-room load spreadsheet, answer a question of quantity. They never answer the question of location: where does the cold air actually arrive, at what temperature does it enter that specific rack, and what becomes of that distribution when one unit stops. Only a spatially resolved solution of the flows answers those three questions.

On eolios.eu Understanding how a data center works

A frequent mistake is to treat the data center as a single object. In reality, the heat crosses five successive thermal barriers, each with its own physics, its own resistances and its own failure modes. A CFD study only makes sense if you know at which scale the question sits.

The five successive thermal barriers, from the millimetre to the site: at each scale, one dominant physics, one critical quantity and one specific failure mode.

The scales are not independent: they propagate. Plume recirculation on the roof (scale 5) raises the air temperature at the dry cooler intakes, therefore the chilled water temperature, therefore the supply temperature in the room (scale 3), therefore the rack intake temperature (scale 2) and finally the junction temperature of the processors (scale 1). A defect at scale 5 is read at scale 1: and conversely, densification at scale 1 ends up visible on the roof.

Scale 3 · data hallComplete modeling of a data hallThe usable volume, the racks and the air handling units reconstructed in a single workable model.

Scale 4 · technical roomsThermals of a technical roomUPS and switchgear rooms: high density per square metre, small volume, a non-negligible share of radiation.

This reading by scales also dictates the modeling scope. Simulating a whole room with heat sink detail would be absurd: the rack is represented by an equivalent model (airflow, ΔT, pressure drops). Conversely, studying an internal hot spot while modeling only the roof makes no sense. The right reflex is to choose the coarsest scale that still contains the cause of the problem, then go down one level if necessary.

On eolios.eu Thermal study of technical rooms (UPS, switchgear)

A thermal study only has value when related to an enforceable criterion. Without a reference framework, a temperature map is an aesthetic object; with one, it becomes proof of compliance or a record of deviation. Here are the texts that structure the practice.

ASHRAE TC 9.9: the reference for intake conditions

The Thermal Guidelines for Data Processing Environments from ASHRAE technical committee 9.9 define, for the air entering the equipment (and not for room air), a recommended range and allowable ranges per equipment class. The usual recommended range is 18 to 27 °C, with an upper relative humidity limit and a dew point envelope. The allowable ranges then extend by class: about 15–32 °C (A1), 10–35 °C (A2), 5–40 °C (A3) and 5–45 °C (A4). The committee has moreover extended its recommendations to liquid-cooled equipment to cover densities above 40 kW per rack, acknowledging that the historical air-centred framework was no longer sufficient.

EN 50600: the European facility standard

The EN 50600 series covers the design, construction and operation of data centre facilities and infrastructures: availability classes, physical protection, power distribution, environmental control and efficiency indicators (including PUE, also standardised at ISO/IEC level). It provides the contractual vocabulary that matters when a study has to prove a level of service, and not only thermal comfort.

Uptime Institute: the redundancy Tiers

The Tier I to Tier IV classification does not speak of temperature but of topology and maintainability. It concerns CFD directly because it defines the configurations to be proven: if the facility claims concurrent maintainability, then thermal behaviour must remain acceptable during the deliberate unavailability of a component. That is not demonstrated at nominal duty, but in a degraded scenario (chapter 14).

Environmental and safety framework: the European scale, the national variants

The structuring requirements are today European; it is their national transpositions that vary from one country to another. The same project run in France, Germany, the Netherlands or Ireland answers the same physical objectives, but before different authorities and permitting procedures.

• F-Gas Regulation (EU) The progressive phase-down of high global warming potential refrigerants steers refrigeration architecture choices and pushes towards free cooling, therefore higher setpoints, therefore a heightened requirement on air distribution.
• Energy Efficiency Directive (EU) Data centres above a given installed power threshold report their performance indicators, including PUE and heat reuse. Thermal proof becomes an argument in the file, not just operational comfort.
• Environmental permitting Generator sets and fuel storage, refrigerants, cooling towers: depending on capacities and quantities, the site falls under a permitting or notification regime, with requirements on exhaust dispersion and health risk control. The ICPE nomenclature in France, the BImSchG regime in Germany or the Dutch environmental permit all pursue the same objective: demonstrating dispersion.
• Cooling towers & legionella When wet cooling is selected, legionella risk control requires a risk analysis and a demonstration of plume dispersion, whatever the jurisdiction.
• Fire safety The classification of the building determines the applicable framework and the smoke control objectives: national regulation of the host country, complemented in practice by the Eurocodes for the structure and by insurer standards (VdS, FM, CNPP) depending on the client. Automatic gas extinguishing has its own design rules (chapter 16).

On eolios.eu Cooling towers & environmental permitting Legionellosis & cooling towers

In the field, thermal disorders almost always come down to a small number of recurring mechanisms. Naming them precisely is the first step of the diagnosis: each has a measurable signature and a distinct remedy. Confusing them leads to treating a symptom with the wrong lever: the textbook case being added cooling capacity to compensate for a sealing defect.

These eight mechanisms combine, and that is what makes diagnosis difficult without simulation: the same temperature measured at the top of a rack may result from recirculation, from a bypass that has emptied the plenum, or from an interaction between two units ten metres away. The probes say that there is a problem; the computed field says where the air goes, therefore which remedy will work.

Pathology 1Recirculation effectsRejected hot air rises back into the cold aisle: the signature is a marked vertical gradient at the rack face.

Project · DC10Distribution of a real data hallMapping of supply and trajectories on one of our instrumented data halls.

On eolios.eu Origin of overheating in data centers

To compare a before and an after, or two design variants, you need indicators that condense a three-dimensional field into a few numbers that can be discussed in a meeting. The industry has standardised a small set of them, all built on the same principle: comparing real thermals with ideal thermals.

The value of these indicators is twofold. In diagnosis, they point immediately to the right family of remedies: an RTI of 70 % with correct intakes signals excess airflow and bypass, therefore a saving potential, not a risk. In design, they make it possible to set performance commitments verifiable in the model, then at handover.

Airflow, the only truly sizing variable

Everything starts from an elementary enthalpy balance. The power removed by an air stream is written:

This relation explains most of the trade-offs of the trade. Increasing the ΔT proportionally reduces the airflow, therefore the fan energy (which varies roughly with the cube of the airflow): it is the most powerful saving lever. But a high ΔT means hotter return air, therefore far heavier consequences in the event of recirculation. Containment is not a comfort option: it is what makes a high ΔT usable without risk.

Working out an order of magnitude

The tool below applies this relation, then derives from it the second figure every operator should know: the rate of temperature rise of the room in the event of total loss of cooling, obtained by dividing the dissipated power by the thermal capacity of the air volume.

On eolios.eu Guide to calculating the PUE of a data center

The choice of a cooling architecture is driven first by density per rack, then by site constraints (clear height, available water, climate, regulation) and finally by operations. CFD comes in at every level, but with different questions: in air, we look for where the air goes; in liquid, we look for what is left for the air to do, because part of the heat continues to be dissipated in the room.

Liquid cooling is progressing fast wherever accelerator loads demand it: some recent computing architectures reach cabinet powers of the order of 120 kW, out of reach of air cooling at full load. The installed base, however, remains overwhelmingly air cooled: the real challenge of the decade is less "air or liquid" than the coexistence of both in the same rooms, with different temperature ranges and different redundancies.

On eolios.eu Cooling systems for data centers Electronics cooling: from the component to the cold plate

CFD (Computational Fluid Dynamics) numerically solves the conservation equations of fluid mechanics over a domain divided into elementary volumes. It "predicts" nothing: it computes the consequences of the assumptions it is given. That is why the quality of a study rests first on the input data and on the choice of models.

The equations solved

In a data hall, air is treated as a weakly compressible Newtonian fluid. Three families of equations are solved simultaneously on each control volume: conservation of mass (continuity), conservation of momentum (Navier-Stokes, including the buoyancy term) and conservation of energy (enthalpy transport, with conduction and possibly radiation). To these are added, depending on the case, transport equations for turbulence, humidity, chemical species (smoke, extinguishing gas) or particles.

The solution is iterative: starting from an initial field, the solver corrects pressure and velocities until the local imbalances (the residuals) become negligible. A distinction is made between steady-state computations (seeking the equilibrium state, the nominal duty case) and transient ones (following the evolution in time), the latter being indispensable for a failure or a fire.

What CFD brings that no other method brings

On eolios.eu Dossier: what is CFD simulation?

A data hall is a case of mixed convection: supply jets impose a forced dynamic, while hot plumes impose a natural one. The two compete for the field, and depending on the balance of forces, the same room can behave very differently. The choice of models is therefore not a software detail: it is an engineering act that must be justified in the report.

Turbulence models

Turbulence is not solved directly: that would require meshes out of reach. Averaged equations (RANS) are solved and the effect of turbulence on the mean field is modeled.

Buoyancy: Boussinesq or variable density

At the moderate temperature differences of a data hall (10 to 20 K), the Boussinesq approximation (constant density except in the gravity term) is legitimate and numerically comfortable. It ceases to be so as soon as the differences become large: dry cooler plumes, generator exhausts, and of course fire, where density must be treated as fully variable with temperature. Using Boussinesq on a fire case is a classic modeling error.

Porous media: representing without meshing

It would be absurd to mesh every tile perforation, every server and every grille. They are represented by equivalent porous media, characterised by a pressure drop law (viscous and inertial Darcy-Forchheimer terms) calibrated on manufacturer data or on measurements. This approach applies to perforated tiles (perforation ratio, presence of a damper), to rack faces, to filters and coils and to return grilles. The quality of that calibration directly determines the fidelity of the airflow distribution: it is often there, and not in the turbulence model, that accuracy is lost.

Modeling racks and air handling units

• Racks Imposed-airflow model (the server fan regulates) or imposed-ΔT model (the power is known). The choice changes the behaviour in a degraded scenario: an imposed-airflow model does not "see" the runaway of internal fans when the incoming air heats up.
• CRAC/CRAH Fan characteristic curve rather than fixed airflow, otherwise the redistribution of airflow when one unit stops and the network changes operating point is ignored.
• Leaks Cable openings, floor joints, unblanked positions. A real room leaks; a perfectly sealed model is systematically optimistic.

The software used in the data center world

The landscape splits into two families. On one side the data center dedicated tools, which embed object libraries (racks, CRAH, perforated tiles, containments) and directly compute the sector indicators. On the other the general-purpose solvers, heavier to deploy but with no limit of scope: this is the only possible choice as soon as a roof, a plume, a fire or an off-catalogue geometry has to be handled.

Radiation

Often negligible between racks, radiation becomes important again in three situations: compact technical rooms with high density per square metre (UPS, switchgear), solar gains on the envelope and roof in external studies, and fire, where it is a major transfer mode and governs spread to neighbouring racks.

On eolios.eu What is a data center digital twin?

Simulation amounts to solving a system of non-linear partial differential equations over a finite number of cells. The mesh is that division. It sets the spatial resolution of the information: a phenomenon smaller than the cell does not exist in the result. On a data hall, that means a coarse mesh will smooth out precisely what is being looked for: intake gradients at the top of racks and local leaks.

Meshing strategy for a data hall

• Hybrid mesh Hexahedra dominant in regular volumes (aisles, plenum), tetrahedra, prisms and pyramids to connect complex geometries. This combines the accuracy of structured meshes with the flexibility of unstructured ones.
• Targeted refinement Rack faces, perforated tiles, supply and return outlets, containment openings, plenum singularities. These are the high-gradient zones.
• Prism layers Near walls, to correctly capture boundary layers and heat exchanges when they matter.
• Geometric quality Skewness, aspect ratio and size jumps between neighbouring cells are monitored. A degenerate cell degrades convergence far beyond its neighbourhood.

Mesh independence, non-negotiable

A result only has value if it no longer depends significantly on mesh fineness. The demonstration is simple and must appear in the report: the same case is solved on at least three meshes of increasing fineness (typically a factor of 1.5 to 2 on the characteristic size) and a few representative output quantities are tracked: maximum intake temperature, RCI, airflow of a critical tile. When the difference between two levels drops below a threshold (often a few tenths of a degree), the mesh is deemed sufficient.

Boundary conditions mathematically express the interaction between the computed domain and its environment. They carry most of the physical truth of the site: it is through them that the simulation learns what the installation really does. Formally, either a value is imposed on the boundary (Dirichlet condition: temperature, velocity), or a flux or gradient (Neumann condition: dissipated power, adiabatic wall), or a mixed relation (heat transfer coefficient, fan curve).

The choice of climatic scenarios deserves particular attention. An installation is sized on unfavourable situations: high-percentile heatwave (not a monthly average), wind direction penalising for plume recirculation, winter for condensation and free cooling questions. In practice, a reduced set of sizing operating points is retained, each justified by climatic data from the most representative weather station.

A simulation always produces images. Nothing guarantees that they describe anything. The distinction is structuring: verification asks "are we solving the equations correctly?", validation asks "are we solving the right equations, with the right data?". Both are necessary and neither replaces the other.

Convergence criteria

Validation: confronting reality

On an existing installation, validation consists in comparing the computation with independent measurements: intake temperatures recorded at several heights, tile airflows, return temperatures, possibly smoke tests to check trajectories qualitatively. The least well known parameters are then adjusted (permeabilities, leaks, real distribution of powers) until an acceptable agreement is reached. That is calibration. It turns a generic model into a digital twin of the site, reusable for all subsequent studies.

Sensitivity analysis and margins

Rather than a single figure, a robust study delivers a range. The doubtful parameters are varied within their credible bounds (IT power ±10 %, plenum leaks, damper opening ratios, outdoor temperature) and the effect on the indicators is observed. This answers the only question that really counts in a project meeting: how much margin do we have before the conclusion changes?

On an existing installation, the most profitable phase of a study is not the computation: it is the measurement campaign. It sets the upper bound of accuracy for everything that follows, and on its own it reveals part of the disorders.

The minimum data set

• Real powers Readings from PDUs or switchboard feeders, rack by rack where possible. Never nameplate powers, which commonly overestimate by 30 to 100 %.
• Intake temperatures At the bottom, middle and top of the rack, on a representative sample; it is the vertical gradient that signs recirculation, not the average value.
• Supply and return temperatures Of each unit, with the setpoint and the fan load factor.
• Perforated tile airflows With an airflow hood: this is the most revealing measurement of an unbalanced plenum, and the one most often missing from the file.
• Plenum pressure At several points, and the real clear height under the floor.
• As-found geometry Up-to-date drawings, real rack positions, empty locations, blanking panels present or not, cable trays, brush grommets.
• Infrared thermography Quick mapping of rack faces, containment doors and technical rooms; an excellent tool for spotting leaks.
• Smoke test Occasional: direct visualisation of trajectories, unbeatable for convincing and for qualitatively validating the model.
• Instantaneous PUE And BMS history: to size the saving potential and have a point of comparison after action.

Qualitative validationSmoke test on a load bankSmoke makes trajectories visible: it is the most convincing tool for confronting a computation with reality.

After calibrationThe digital twin of the siteOnce tuned on measurements, the model becomes reusable at every change of load.

On eolios.eu Data center audit & diagnostics

At nominal duty, the objective is not to produce a beautiful map but to answer a finite list of questions, in an order that goes from the global to the local. This sequence is our standard analysis grid.

The sixth question is the one that produces the most economic value, and it is the one that cannot be asked without simulation. Raising the supply setpoint by 2 to 3 °C reduces cooling production consumption and considerably extends the annual hours of free cooling: but only if there is proof that no rack leaves its allowable range, including at the top of the worst-placed rack, including when one unit is stopped. CFD provides exactly that proof.

Project · rooms 5 & 6Two data halls in parallelComparison of two neighbouring rooms: same design, different airflow behaviour.

ASHRAE criterionInlet temperatures, rack by rackThe only criterion that counts: the air at the intake, over the whole usable height, not the room average.

Technical rooms, a classic blind spot

Server rooms concentrate attention; UPS, switchgear and transformer rooms concentrate incidents. They combine high density per square metre, a small volume, often rudimentary ventilation, a share of radiative transfer and strict manufacturer thresholds on batteries. They are excellent candidates for a dedicated study, generally quick and with a high return.

On eolios.eu Data center CFD engineering Technical rooms

A well-designed data center is not judged at nominal duty, where almost everything works, but on its ability to get through an incident. This is the field where CFD has no competitor: nobody agrees to shut down a chiller in production to see what happens.

The scenarios to be investigated

• Loss of one air handling unit The total airflow drops, but above all the distribution map changes. Neighbouring units redistribute according to their curve, and the service zone lost is not the one you would expect. Some tiles may reverse.
• N+1 configuration under maintenance Identical to the previous one but deliberate and prolonged; it is the proof required by concurrent maintainability claims.
• Loss of a chiller or of a loop The supply temperature drifts while the airflow remains nominal. The signature is very different from a loss of airflow.
• Power cut and switch to backup For a few seconds to a few tens of seconds, the unit fans stop while the IT equipment, on UPS, keeps dissipating. The plenum depressurises very fast, the room moves into a pure buoyancy regime.
• Prolonged heatwave Cooling production loses efficiency and free cooling is no longer available; the internal study must then be coupled with the external one (chapter 15).
• Containment breach during works Open door, panels removed, rack being replaced. A mundane scenario, rarely investigated, often responsible for alarms.

These orders of magnitude are enough to understand why the subject has changed nature with density: at 200 kW in a large volume, the operator has time to react; at 800 kW, the automatic fan restart sequence becomes the only safeguard, and its response time must be compared with the time to threshold exceedance. That is precisely what a transient simulation provides: not an average, but the intake temperature curve of each rack, second by second.

The whole cooling chain depends on a single physical quantity: the temperature of the air actually drawn in by the outdoor heat exchangers. And that temperature is not the one from the weather station. It results from the interaction between the hot exhausts of the installation, the geometry of the roof and of the site, and the wind. That is the object of external CFD.

Plume recirculation, the central mechanism

Plume recirculation is a self-aggravating mechanism: the machine whose inlet air heats up sees its performance fall, draws more power, therefore rejects more heat, and feeds its own plume. The at-risk configurations are well identified: machines too close together, exhausts directed towards a neighbouring intake, high parapets trapping hot air, badly sized acoustic enclosures, and neighbouring buildings creating a wake that pushes the plumes back down onto the roof.

The questions handled externally

• Temperature rise at the intake Of each machine, by wind direction and speed, under heatwave conditions.
• Rooftop layout Spacing, discharge height, orientation, ducting or discharge nozzle, screens and windbreaks.
• Generator sets Dispersion of hot exhaust gases, interaction with fresh air intakes and dry coolers, a scenario where both systems operate simultaneously.
• Pressure drops Induced by enclosures, grilles, louvres and rooftop photovoltaic panels.
• Heat island Impact of the heat rejection at neighbourhood or campus scale, a subject that has become central in permitting procedures.
• Air quality and acceptability Plume dispersion, a health question for wet cooling devices.

Heat rejectionRooftop plumes, second variantSame roof, different layout: the path of the plume completely changes the intake temperature.

OptimizationOptimization of thermal plumesDischarge height, orientation, screens: every degree gained at the intake is found again in the cooling production.

The link with the internal study is direct and quantifiable: every degree gained at the machine intakes translates into available cooling capacity, therefore chilled water temperature, therefore margin on the supply setpoint. That is why we increasingly handle both scopes in a single coupled assignment, rather than as separate studies.

Heat islandCampus of 24 data centersLarge-scale study: the cumulated heat rejection of a whole campus and its effect on the ambient air.

Wind & siteExternal wind studyThe wake of neighbouring buildings pushes plumes back down: the wind decides the temperature seen by the machines.

On eolios.eu External CFD simulation of a data center Heat island impact

The fire risk of a data center has its own characteristics: a high concentration of electrical energy, combustibles producing dense, toxic and corrosive smoke, and a stake that is not only human but also material: combustion products attack electronics far beyond the burned zone. Simulation changes objective here: we no longer look for comfort, we look for tenability and control.

What is computed

• Fire development And possible spread to neighbouring racks, radiation included.
• Temperature, visibility and toxicity fields At head height, along the evacuation and intervention routes.
• Effectiveness of smoke control Extraction airflows, position of vents and air inlets, smoke-free height, stagnation zones.
• Interaction with comfort ventilation An AHU that keeps running can stir the smoke and destroy the stratification that smoke control seeks to establish.
• Automatic gas extinguishing Dispersion and homogeneity of the agent concentration, hold time of the concentration, aeraulic effects of the discharge and overpressure.
• Compartmentation Real effectiveness of the separations and of the closing devices.

Two modeling requirements set fire apart from the rest: the computation is necessarily transient, over durations of several tens of minutes, and density must be treated as fully variable: the Boussinesq approximation has no validity left with differences of several hundred degrees. Radiation becomes a first-order transfer mode, and the definition of the design fire (power, growth curve, area, soot production) must be justified against the applicable framework.

Smoke controlIsosurfaces of fresh and hot airThe stratification that smoke control seeks to establish: and that comfort ventilation can destroy.

Gas extinguishingDispersion of NOVEC agentHomogeneity of concentration and hold time: the two criteria of an effective gas extinguishing system.

On eolios.eu Data center fire simulation Automatic gas extinguishing systems

PUE relates the total energy of the site to the useful energy of the IT equipment. It is simple, imperfect, universally used. And it is stagnating: 1.54 as a global average, unchanged for six consecutive years according to the Uptime Institute, against 1.10 to 1.15 claimed by hyperscalers and 1.58 to 1.80 observed in colocation and enterprise. The gap is not technological: it is aeraulic and operational.

The five levers simulation acts on

Conversely, one has to be clear about what CFD does not do: it improves neither the efficiency of a chiller, nor that of a UPS, nor the load factor of the IT equipment. A PUE degraded by a very low load factor is not corrected by an airflow study: it is corrected by consolidation. Knowing how to tell the two situations apart avoids selling, and buying, the wrong study.

On eolios.eu Guide to calculating PUE Energy optimization & PUE

Our reference commitment on a data hall is four weeks, from scoping to presentation: and we know how to go faster when the project schedule demands it.

A CFD study is not an isolated computation but a sequence of decisions shared with the design team and the operator. The sequence below is the one we apply: one calendar month, with phases partly run in parallel. Our in-house computing resources make it possible to compress that lead time to two or three weeks in an accelerated procedure: a commissioning trade-off, an urgent layout decision, an incident to investigate. Only two situations really extend it: a very large number of transient scenarios, and waiting for input data on the operator's side.

Six assignments, six different mechanisms. The detailed case files, with configurations and results, are published on this site.

Project · DC25Plume synthesis, 27 °CBehaviour of the heat rejection at a high setpoint: the case where the usable limit is being sought.

Project · hyperscaleExternal study of a hyperscale siteEnvelope and roof of a high-density site, handled together with the room in a single assignment.

Project · ItalyData hall & UPS roomsElectrical rooms often turn out to be more constrained than the server room itself.

Project · backupPressure drops of a generator setEnclosures, grilles and louvres: what the silencer really costs in airflow terms.

On eolios.eu All our data center projects

These situations come back on almost every site we audit. None of them stems from a sophisticated engineering error: they are operational drifts or design shortcuts, with a disproportionate thermal effect.

An honest white paper must say where the tool it presents stops. CFD is powerful; it is also easy to misuse, and a wrong result is visually indistinguishable from a right one.

What CFD cannot do

• Compensate for bad input data. An approximate IT power or an obsolete geometry produces an approximate result, whatever the refinement of the computation.
• Predict the behaviour of a control system that is not modeled. Controllers react to measurements: if the control strategy is not described, the computed transient is only one scenario among others.
• Replace measurement. It organises and interprets measurement; it does not replace it, and on an existing site it depends on it.
• Settle non-aeraulic questions. Electrical reliability, cybersecurity, machine efficiency: out of scope.
• Stay valid indefinitely. A model describes one state of an installation. Any change of load or layout calls for an update of the digital twin.

Eight questions to ask before ordering a study