Acoustic & Emissions Engineering Design

This is where the siting and permitting chapters turn into steel, catalyst and concrete. A number written into a permit condition two years earlier either gets met, or the project gets a stop-work order, a consent decree or a midnight call from a county supervisor. The design targets here are externally imposed thresholds, not engineering optima: the property-line dBA/dBC limit in the noise ordinance (Chapter 3.11), and the ground-level NO₂/CO/PM concentrations and stack parameters in the air permit (Chapter 3.9). Source emissions are written in NOx, but the US ambient standard is for NO₂, so the dispersion demonstration converts one to the other with a regulator-accepted method and background concentration. The engineer does not choose how quiet or how clean to be. The job is to hit someone else's number at the lowest capital and footprint cost, on equipment whose noise and emissions follow from power and cooling decisions made elsewhere in Part 6.

The first variable to price is the one nobody wants to spend: land. Doubling the far-field distance from an unchanged point source buys roughly 6 dB of geometrical spreading, with no attenuation hardware to maintain. Ground, weather, reflections and other sources still shape the level at the receptor. Price setback against enclosures, attenuators and low-frequency source treatment at that receptor, because a costly parcel can reverse the apparent saving even when distance lowers the level. Emissions follow the same logic. The cheapest NOx is the NOx you never burn (grid power, fuel cells), and after that comes the SCR you sized into the generation island at design time instead of retrofitting under a consent decree. Four physical systems do the work: acoustic enclosure and attenuation, emissions control, stack and dispersion, and vibration isolation.

Before you select equipment, the design basis must capture two externally fixed numbers and one nuance. The first is the property-line sound limit, usually a day band and a stricter night band. Prince William and Stafford counties in Northern Virginia enforce 60 dBA day / 55 dBA night at residential receivers (Fairfax County’s April 2025 ordinance survey), and some jurisdictions collapse the two into a single 24-hour limit. Check the zoning class and the weighting before you assume either. Prince William’s Ordinance 25-62, adopted 28 October 2025 and enforced from 1 May 2026, keeps 60/55 dBA at residential receivers and adds a residential steady-tonal limit of 73 dBC day / 68 dBC night over a five-minute interval. Its industrial-district limits remain 79/72 dBA, and its octave-band values are advisory.

Virginia added a statewide process in 2026. Under HB 153 (Code §15.2-2209.5, signed in April and in force since 1 July 2026), a new facility expected to need 100 MW or more from its retail electric utility must have a sound-profile assessment of residences and schools within 500 feet of the property boundary before the locality approves a rezoning, special exception or special use. Expansions of existing approved sites that add less than 100 MW are excluded. The locality checks the assessment against its own noise ordinance; the law itself sets no sound limit and no setback.

The second number is the air-permit limit: the allowable mass-emission rates, and the ground-level concentration limits that a dispersion model must show you meet. The nuance, and the one that wrecks projects designed to the old rules, is the move from A-weighting to C-weighting and octave-band criteria.

A-weighting deliberately discounts low-frequency energy, because the ear is less sensitive to it at moderate levels. But the dominant annoyance from a data center is low-frequency: the tonal hum of large axial fans, dry-coolers and transformers. This guide uses 20–200 Hz as its working band for that hum. Leventhall’s general definition of low-frequency noise is roughly 10–200 Hz, and the published data-center cases with spectral detail report tones near 70–140 Hz (Bray & Dominici, 2026 narrative review). The hum travels farther than mid-band noise, because low frequencies attenuate less with distance and pass through barriers and walls that stop high frequencies. The A-filter then subtracts it from the reported number: ASHRAE’s A-weighting correction at 63 Hz is 26.2 dB of attenuation, so a 75 dB band level there contributes under 49 dBA before it is summed with the rest of the spectrum. The tone is measurable; the metric hides it. A facility can pass a 55 dBA night limit at the fence and still draw complaints, because the neighbors are hearing a 63 Hz tone the A-filter removed. The 2026 ordinance wave answers with C-weighted (dBC) limits, tonal penalties and, in some jurisdictions, advisory or binding octave-band criteria that the hum cannot hide from. If your jurisdiction has made that move, an A-weighted-only acoustic model is not conservative. It is wrong, and it will certify a design that fails the actual test.

The mechanical and electrical plant is a stack of point and line sources, each with its own spectrum. Gensets and prime-power engines are broadband and loud: casing, exhaust, radiator air and intake all radiate. Where an ordinance allows night operation or testing, they can be the worst-case audit, because night limits are the tightest (Prince William limits routine exercising to the daytime). Chillers and dry-coolers are the steady, always-on sources, and their large low-speed fans produce the dominant low-frequency tone. Cooling towers add water-fall and fan noise. Air-cooled chillers and the fan walls of a liquid-to-air heat-rejection plant are the modern density driver: as the heat load per hall climbs with the rack ramp (Chapter 5.8), installed fan power and radiated acoustic power climb with it.

Compare the mitigation toolkit by receptor spectrum, land cost and the extra power or maintenance each measure needs. Setback is the first layout option to price: it buys geometrical attenuation without hardware where the parcel and source geometry allow it. Source selection comes next. EC (electronically commutated) motors help efficiency and part-load speed control, but quiet comes from the fan’s aerodynamics. Ask each vendor for sound-power spectra at the same airflow and static pressure, across the speed range. Choosing a fan whose blade-pass frequency avoids a resonance can move the tone out of the worst annoyance band. Attenuators (parallel-baffle silencers on intake and discharge) and acoustic louvers need a tested insertion-loss spectrum and cost static pressure. The fans must overcome that with more power, partly re-radiated as noise, and the cooling plant must budget for it. Barriers and berms depend on geometry and spectrum: FHWA’s highway guidance treats about 5 dBA as the result when line of sight is just blocked, 10 dBA as attainable and 15 dBA as very difficult. Sound diffracts around their edges, and useful low-frequency reduction needs the right geometry, depth and a modeled band-specific result for your source and receiver. Full enclosures are the last and most expensive resort, and they create their own ventilation and heat-rejection problem.

The table is a sequencing rule. Compare land, source selection and hardware at the same receptor before spending on any of them. Every row below setback either costs static pressure (which the fans pay back partly as noise and the cooling plant pays as energy), creates a heat-rejection problem of its own (enclosures), or fails on the band that matters (barriers at low frequency). A costly mistake is to skimp on setback at scoping and buy the smaller, cheaper parcel. Then commissioning shows that no achievable mix of attenuators and barriers makes up the night-time low-frequency gap. The options left then include operational curtailment (running the plant below capacity at night, which strands the power you came for), buying adjacent land at a distressed-buyer premium, or source retrofits whose cost depends on the equipment. The acoustic budget is set on the site plan, years before the first measurement.

Calculate bands before summing. Per source, Lp = Lw + 10 log10(Q) − 20 log10(r/1 m) − 11. Add 10 log10(8) for eight independent sources; Q = 2 and r = 100 m give Lw − 38.9588 dB. Subtract each assumed band reduction, add its A-weighting correction, then compute 10 log10 Σ10^(adjusted band/10). The plant level falls from 54.7705 to 47.7448 dBA, displayed as 55 to 48 dBA. Add background energetically: 10 log10[10^(47.7448/10) + 10^(45/10)] = 49.5960 dBA. Then add the tonal 3 dB and the 2 dB allowance: 54.5960 dBA, about 55 dBA.

Take the mitigated cooling package forward to source testing; the unmitigated one fails. Compare unrounded values: 54.5960 leaves only about 0.4 dB below the 55 dBA night limit. Before adjustments, the plant-plus-background crossover is 50 dBA, which allows about 48 dBA from the plant alone. A small attenuation shortfall reverses the result, and rounding down cannot establish a pass.

Routine generator testing changes the answer. Add the two assumed 45 dBA generators. Total = 10 log10[10^(47.7448/10) + 3 × 10^(45/10)] = 51.8855 dBA before assessment adjustments; adding 3 + 2 dB gives about 57 dBA. That fails the assumed 55 dBA night condition and passes the 60 dBA day condition. Test simultaneously in the daytime under the actual permit; emergency operation has separate conditions. In Prince William, routine exercising of emergency generators is allowed only in the daytime, and the extra steady-tonal limits are waived only during that exercising, so attenuation cannot buy night exercising there. Using unrounded model intermediates, the simultaneous-test crossover is a combined generator contribution of 10 log10[10^((55−3−2)/10) − 10^(47.7448/10) − 10^(45/10)], about 39 dBA. The assumed pair is about 48 dBA, so where an ordinance or permit did allow night testing, that pair would need roughly 9 dB more attenuation at the receptor. A quieter pair reverses the decision only after the unrounded total and the separate tonal and band checks pass.

The field handoff specifies receptor position and height, equipment count and speed/load, averaging time, meteorology, calibrated meter and background-on/off method. It also states which tonal method the permit uses. Log dBA and dBC, the unweighted third-octave bands, and a narrowband spectrum that can separate the suspected tones. Judge acceptance by that permit method. If you also screen with Annex K of ISO 1996-2:2017, which is informative, know that its band-to-neighbor level differences are offered as possible choices. Arup’s Strutt software, for example, tests each band against both neighbors at 5 dB (500–10,000 Hz), 8 dB (160–400 Hz) and 15 dB (25–125 Hz). A screen that finds nothing does not rule out an audible tone. Apply the permit’s tonal adjustment in place of the worked case’s 3 dB. The worked case also ignores coherent fan tones and building and ground reflections, so release the design only after installed-mode measurements pass, including the low-frequency conditions. ASHRAE’s Sound and Vibration chapter supplies summation, weighting and propagation principles; Chapter 3.11 covers the enforceable community commitments.

The emissions problem arrived with the same force as the noise problem, from the same root cause: on-site generation. When the grid could not energize the load on an AI timeline, operators brought their own power, in reciprocating engines and aeroderivative or industrial gas turbines. That turned a passive electrical load into a stationary combustion source with a Clean Air Act footprint (the permitting side lives in Chapter 3.9; the fuel-process and gas-handling engineering in Chapter 4.9). This chapter owns the engineering response: the post-combustion catalyst train that takes raw exhaust under the permit limits.

The two workhorses are Selective Catalytic Reduction (SCR) for NOx and an oxidation catalyst for CO and unburned hydrocarbons (VOC). SCR injects a reductant, aqueous urea or anhydrous/aqueous ammonia, upstream of a catalyst bed, where NOx is reduced to nitrogen and water vapor. For affected new natural-gas turbines, EPA’s subpart KKKKa (final January 2026, corrected July 2026) offers an input-based concentration option, measured on a four-operating-hour rolling average at 15% O₂; an output-based route is also allowed. For units above 850 MMBtu/h at utilization above 45% the option is 5 ppmvd NOx. Units of 50–850 MMBtu/h get 15 ppmvd at utilization above 45% and 25 ppmvd at 45% or below. A BACT or LAER determination and the site permit can go lower, and the permit may require dry low-NOx (DLN/DLE) combustion with SCR. The oxidation catalyst treats the CO and hydrocarbons that combustion leaves, and its temperature and oxygen window must suit the turbine or engine exhaust. None of this is free. SCR adds backpressure (a turbine derate), needs a reductant storage-and-dosing system (an ammonia inventory with its own process-safety and setback implications), requires the exhaust to sit in a catalyst-friendly temperature window, and leaks a small ammonia slip that is itself a permit-limited pollutant. Get the temperature window or the dosing control wrong and you trade a NOx exceedance for an ammonia-slip exceedance.

Separate the machines before specifying catalysts. A lean-burn reciprocating engine can use SCR for NOx with suitable oxidation control, and a rich-burn engine can use nonselective catalytic reduction with a controlled air/fuel ratio. EPA’s October 2024 AP-42 natural-gas engine chapter describes those combustion and control differences. KKKKa turbine concentration limits do not become RICE guarantees. Get rates for normal load, turndown, startup and catalyst-unavailable states, plus the enforceable hours and start counts, before the air permit is final.

Diesel standby generators are a third case. NSPS Subpart IIII sets the rules. A current standby diesel above 560 kW and under 10 L per cylinder (not a fire pump) only needs Tier 2 certification when it is an emergency unit. A new engine in ordinary non-emergency duty must meet the Tier 4 standards that apply to it. Specify the certified engine and control package rather than assuming SCR plus a particulate filter. Maintenance and readiness testing may use up to 100 h/yr, and up to 50 h of that same allowance may be non-emergency operation. Federal rules put no hour limit on real emergencies. Running for paid grid support must satisfy each dispatch condition in §60.4211(f)(3)(i). A general demand-response allowance is not enough. Chapter 3.9 covers applicability. The European comparison is narrower than it looks. The Medium Combustion Plant Directive offers an optional exemption from emission limit values for plants that run 500 h/yr or less, averaged over five years for existing plants and three for new ones. That is not an operating cap, and site permits set their own. The Environment Agency’s Equinix Slough decision shows one version: 500 hours of emergency use for the whole installation, counted as elapsed time whenever any generator is running, with testing limited separately. Schedule the test regime against the hour limits and the acoustic night limit together.

An air permit is granted against ground-level concentrations at and beyond the property line, not mass-emission rates alone, and a dispersion model (AERMOD in the US, or the regional equivalent) demonstrates them. The model is parameterized by stack attributes the engineer controls: stack height, inside diameter, exit temperature and exit velocity. Plume rise, how high the hot, fast exhaust climbs before it dilutes, depends strongly on buoyancy (temperature) and momentum (velocity). A stack that is too short, too wide (low velocity) or too cool gives up plume rise, and the modeled ground-level concentration at the fence rises with it. The subtle part is building downwash. A stack that does not clear the aerodynamic wake of the data hall (the 'good engineering practice' stack-height rule) has its plume dragged into the building’s recirculation cavity, spiking concentrations exactly where the receptors are.

The stack you build must match the stack the model assumed. Suppose the permit was modeled on a 30 m stack at 25 m/s exit velocity. If value engineering later shortens it, or the as-built turbine derate drops the exhaust temperature, the source no longer disperses the way the permit certified, and a fully compliant mass-emission rate can still produce a modeled (and real) exceedance at the boundary. Treat the dispersion model’s stack parameters as permit conditions, not design suggestions: changing them is a permit modification, not a field decision. This is also where the rain-cap fight lives. A horizontal or capped discharge protects the catalyst from rain but suppresses vertical momentum, while the plume’s thermal buoyancy remains. AERMOD models capped and horizontal releases explicitly (the POINTCAP and POINTHOR treatments, with building downwash; EPA AERMOD User’s Guide, July 2026, §3.2.2.4), so model the actual release geometry. Where the dispersion demonstration and the permit call for a vertical uncapped discharge, the mechanical design must accommodate it.

Not all of the low-frequency problem travels through the air. The rotating and reciprocating plant (chillers, pumps, dry-coolers, engines) feeds vibration into the slab. The structure re-radiates it as low-frequency sound and transmits it as perceptible vibration into adjacent occupied space and, occasionally, neighboring buildings. No exterior barrier or enclosure interrupts this structure-borne path, because the energy never went through the air. It is the mechanism behind the most stubborn hum complaints: the energy couples into the building at the equipment feet and re-emerges as a tone wherever the structure connects to ground.

The fix is isolation at the source, designed in at equipment set. Use resilient isolators, with an inertia base carried on them where the equipment needs one. A base adds mass and rigidity but does not replace the springs, and neither guarantees that the path disappears. Select stiffness and deflection against the lowest forcing frequency across the operating speeds, so the isolator’s natural frequency sits well below it (the lower it is relative to the disturbance, the more energy is rejected). Put flexible connectors on every pipe and duct penetration, so the charged liquid-cooling loop (Chapter 5.7) does not become a rigid bridge that carries pump vibration across the building. Use seismic-rated isolation that survives the anchoring requirements of Chapter 6.7 without a rigid restraint short-circuiting it. Skip this and you get the worst kind of complaint. An airborne barrier or attenuator cannot cure it, and setback helps only as far as frequency, geology and foundation allow. You find it after the plant is running, when the fix means diagnosis and possibly disruptive isolation, connection or foundation work.

Acoustic and emissions engineering is a backward-solved problem. The targets are fixed externally (the ordinance in Chapter 3.11, the air permit in Chapter 3.9), and the engineer’s degrees of freedom are the placement, hardware and source-selection choices that meet them at the lowest footprint and capex. Three rules follow. Price land against hardware. Setback changes geometrical spreading, while low-noise sources and engineered attenuation can also cut low-frequency sound, so the acoustic outcome is decided on the site plan, not in the silencer catalog. Design the controls in, not on. SCR, oxidation catalyst and vibration isolation cost little as integrated line items and a great deal as retrofits. Temporary status under subpart KKKKa lasts at most 24 consecutive months at one site, and it never exempts a turbine from a catalyst that BACT, LAER or the site permit requires. And match the as-built to the model. The dispersion model’s stack parameters and the acoustic model’s octave-band assumptions are permit conditions, so value-engineering a stack shorter or skipping the dBC analysis is not a saving. It is a re-permit waiting to happen.

Choose a plant and operating schedule that meet each receptor’s actual acoustic conditions and the generation permit’s equipment-specific limits. If the installed mode fails, repair or restrict that mode and retest it. A passing normal cooling run does not buy permission for night generator testing.