# Duct-Borne Noise: Fan to Room

> One page of the phonometry documentation, as Markdown. Index: https://docs.jmrp.io/phonometry/llms.txt

Canonical: https://docs.jmrp.io/phonometry/devices/noise-control/duct-path/
Language: en
Alternate: https://docs.jmrp.io/phonometry/es/devices/noise-control/duct-path/index.md
Version: phonometry 4.0.0
Install: pip install "phonometry[full]"
Author: José Manuel Requena Plens (https://orcid.org/0000-0003-1250-6212)
Published: 2026-07-30
Updated: 2026-11-16T21:36:14Z
Standards: VDI 2081 Blatt 1:2001-07; AHRI Standard 885
Verified: 18 conformance checks (ASHRAE handbook: HVAC applications (SI edition); VDI 2081 Blatt 1:2001-07) https://docs.jmrp.io/phonometry/reference/conformance/
License: MIT (https://github.com/jmrplens/phonometry/blob/main/LICENSE)
Summary: HVAC duct noise from fan to room: fan power, attenuation, regenerated noise, room effect and NC verdict, checked on Long's Table 14.9.

---

Air-conditioning noise is not predicted, it is *accounted for*. You start
from the sound power the fan puts into the duct, walk down the path, and at
every element subtract what it attenuates and add back what its own airflow
regenerates. What survives to the terminal device is turned into a sound
pressure level by the room, the supply and the return paths are added
together, and the total is laid against the design criterion. If it fails,
the sheet itself tells you which element to change: the row with the small
attenuation, or the row whose self-noise is now the floor.

That bookkeeping is what `noise_control.duct_path` implements, with the
element models of `noise_control.hvac` feeding it and
`noise_control.duct_modes` marking the frequency above which the whole
one-dimensional picture stops being exact. The reference throughout is
Long, *Architectural Acoustics* (2nd ed.), Chapters 13 and 14, whose
Table 14.9 is the worked sheet this guide is built around, with the ASHRAE
*HVAC Applications Handbook* Chapter 49 for the air terminal devices and
Bies, Hansen & Howard for the splitter silencers and the plenums. The
reactive four-pole silencers of
[Silencers](https://docs.jmrp.io/phonometry/devices/noise-control/silencers/) and the rest of the installation
methods in
[Industrial noise control](https://docs.jmrp.io/phonometry/devices/noise-control/noise-control/) are the
companion pages.

## 1. The sheet, and how it adds up

A duct-borne calculation is a table: octave bands across the columns
(63 Hz to 8 kHz, the range the published procedures use), one block of rows
per physical element. Each block prints what the element takes out, the
running level after subtracting it, what the element puts back, and the
level leaving it. `DuctElement` carries exactly the two spectra an element
owns, and `duct_path` walks them:

```python
from phonometry import noise_control

bands = [63.0, 125.0, 250.0, 500.0, 1000.0, 2000.0, 4000.0, 8000.0]
fan = [90.0, 86.0, 82.0, 79.0, 77.0, 75.0, 71.0, 61.0]

path = noise_control.duct_path(
    fan, [
        noise_control.DuctElement(label="Elbow, 36 x 24 in, unlined",
                                  attenuation=[0, 1, 2, 3, 3, 3, 3, 3],
                                  self_noise=[41, 39, 36, 29, 20, 6, 0, 0], code="2"),
        noise_control.DuctElement(label="Silencer, 3 ft, standard pressure drop",
                                  attenuation=[7, 12, 16, 28, 35, 35, 28, 17],
                                  self_noise=[49, 43, 44, 42, 42, 45, 35, 24], code="3"),
    ],
    frequencies=bands,
    source_label="Fan, centrifugal FC, 5000 cfm, 2 in w.g.",
)

for row in path.table():
    print(f"{row['code']:>2} {row['label'][:38]:<38} "
          f"{[round(float(v)) for v in row['values']]}")
```

```text
 S Fan, centrifugal FC, 5000 cfm, 2 in w. [90, 86, 82, 79, 77, 75, 71, 61]
 2 Elbow, 36 x 24 in, unlined             [0, -1, -2, -3, -3, -3, -3, -3]
   Sum                                    [90, 85, 80, 76, 74, 72, 68, 58]
   Self-noise                             [41, 39, 36, 29, 20, 6, 0, 0]
   Combined                               [90, 85, 80, 76, 74, 72, 68, 58]
 3 Silencer, 3 ft, standard pressure drop [-7, -12, -16, -28, -35, -35, -28, -17]
   Sum                                    [83, 73, 64, 48, 39, 37, 40, 41]
   Self-noise                             [49, 43, 44, 42, 42, 45, 35, 24]
   Combined                               [83, 73, 64, 49, 44, 46, 41, 41]
Lp Received level                         [83, 73, 64, 49, 44, 46, 41, 41]
```

Three conventions matter, and they are worth stating once because every
published sheet states them differently. All three can be read off the rows
above.

**Attenuations are positive.** Every element model in `noise_control.hvac`
returns a loss as a positive number of decibels, and the cascade subtracts
it. Printed worksheets show the same quantity as a negative level change,
so `DuctPathResult.table()` flips the sign back for the `"attenuation"`
rows: the table reads like the reference, the arithmetic does not have to.
The minus signs on the two element rows above are that flip, not a
subtraction the reader has to make.

**Regenerated noise adds on a power basis.** The self-noise of an element
is a sound power level in its own right, not a correction to the incoming
level, so it is combined as
$10\log_{10}(10^{L/10} + 10^{L_\text{sn}/10})$ rather than added
arithmetically. That is why the `Self-noise` row sits between the `Sum` and
the `Combined` row and never touches the attenuation. Watch the silencer's
triple above: at 63 Hz the running level is 83 dB and the self-noise 49 dB,
so the combination is still 83 dB and the silencer's own noise is invisible;
at 2 kHz the attenuation has done its work, the sum is down to 37 dB against
a self-noise of 45 dB, and the combination comes out at 46 dB: the element
bought to remove the fan is now the loudest thing in that band.

**There is a self-noise floor.** Long's sheet uses a 0 dB sound power level
wherever an element has no regenerated-noise data, and also as a floor
under any computed level that would go negative, which is why his received
spectrum bottoms out near 0 dB instead of running off to minus infinity.
`self_noise_floor` reproduces that (default `0.0`) and `None` switches it
off entirely; it engages in the 4 kHz and 8 kHz rows of the full sheet in
section 6, where the supply path runs past it.

## 2. The source: fan sound power

The fan is the one element whose spectrum you can build from the operating
point alone. `fan_sound_power` implements the ASHRAE scaling law printed as
Long Eq. 13.1,

$$
L_W = K_\mathrm{F} + 10\log_{10}\frac{Q_\mathrm{F}}{Q_\text{REF}}
    + 10\log_{10}\frac{P_\mathrm{F}}{P_\text{REF}} + C_\text{EFF} + C_\text{BFI},
$$

with the spectral constant $K_\mathrm{F}$ of Table 13.5 (one row per fan type), the
off-peak efficiency correction $C_\text{EFF}$ of Table 13.6 and the blade
frequency increment $C_\text{BFI}$ of Table 13.7 dropped into the single
octave band that contains the blade passing frequency. In SI the references
are $Q_\text{REF} = 0.472$ L/s and $P_\text{REF} = 249$ Pa, so the two
logarithmic terms take the same values as the foot-pound form in cfm and
inches of water gauge.

```python
from phonometry import noise_control

CFM, IN_WG = 0.0004719474432, 249.0

fan = noise_control.fan_sound_power(volume_flow=5000 * CFM, fan_static_pressure_pa=2 * IN_WG,
                                    fan_type="forward_curved", relative_efficiency_percent=80.0)
print([round(float(v)) for v in fan.values])
# [99, 99, 89, 84, 82, 77, 72, 67]

print(noise_control.fan_efficiency_correction(relative_efficiency_percent=80.0))          # 6.0 dB off the peak
print(noise_control.blade_passing_frequency(1200.0, 24))      # 480.0 Hz, in the 500 Hz band
print(noise_control.fan_casing_attenuation().values)          # what the housing holds back
# [ 0.  0.  5. 10. 15. 20. 22. 25.]
fan.plot()                                      # the band spectrum, one line
```

The `fan_type` string is the first decision of the whole calculation and the
largest single lever in it, because it selects the row of Table 13.5 that sets
both the level and the shape. The function takes `"airfoil_large"` and
`"airfoil_small"` (backward-curved or backward-inclined centrifugal wheels above
and below 36 in), `"forward_curved"`, three radial rows by total pressure, three
`"vaneaxial_hub_*"` rows by hub ratio, two `"tubeaxial_*"` rows by wheel
diameter, and `"propeller"`. Ranked as a designer ranks them: the airfoil and
backward-inclined centrifugals are the quiet, efficient choice for ducted air
(86/86/88/80/76/69/65/63 dB for this duty), forward-curved wheels are cheap and
compact but much noisier at low frequency (99/99/89/…), and vaneaxial and
propeller fans put their energy into the mid and high bands
(propeller: 99/97/104/102/101/98/92/88 dB). The spectrum shape changes with the
row, not only the level, so the choice decides which silencer will be needed.

$1200$ in `blade_passing_frequency(1200.0, 24)` is the shaft speed in
**revolutions per minute** and 24 the blade count; the increment $C_\text{BFI}$
it produces is dropped whole into the single octave band containing the result.
A fan whose blade tone lands in the band that already governs the design is
worth reselecting rather than silencing. And because that increment represents a
*pure tone* inside an octave band, a sheet that passes on band levels can still
be judged intrusive; that is one of the limits listed at the end of this guide.

Two habits keep this honest. The law assumes ideal inlet and outlet flow
conditions, so a fan boxed into a plant room with a bad inlet is louder
than it says; and ASHRAE's own current guidance is that a fan's sound power
"is best obtained from manufacturers' test data" to AMCA Standard 300 or
ASHRAE Standard 68. Treat Eq. 13.1 as the early-design fallback, not as the
answer. What that test is matters too: **ISO 5136** determines the sound power
a fan radiates *into a duct* with an in-duct microphone carrying a slit sampling
tube behind a turbulence screen, precisely because a bare microphone in a
flowing duct measures its own wind noise long before it measures the fan (the
standard puts the ceiling at 15 m/s for a foam ball, 20 m/s for a nose cone and
40 m/s for the sampling tube). So a fan's duct-borne sound power cannot be
obtained by standing beside the machine with a sound level meter: that measures
the casing-radiated power of `fan_casing_attenuation` instead, which is the
other row of the same source. The fan radiates the same power from its intake
and from its discharge, which is why the supply and return paths of a real sheet
start from the *same* row.

`fan_efficiency_correction` is a step function, and a brutal one: a fan
running at 90 per cent of its peak static efficiency adds nothing, one at
80 per cent adds 6 dB, one below 50 per cent adds 16 dB. Selecting a fan
away from its best point is the cheapest way to lose a duct-noise budget
before any silencer is priced. `fan_casing_attenuation` (Table 13.8) is the
other side of the same source: the power the housing radiates into the
plant room instead of into the duct, zero at 63 and 125 Hz because a
vibrating casing radiates low frequency as freely as the unhoused fan.

![Octave-band sound power of a 5000 cfm forward-curved fan at 2 inches water gauge, from 63 Hz to 8 kHz, at three operating points: 90 per cent of peak static efficiency running from 93 dB down to 61 dB, 80 per cent 6 dB above it and 55 per cent 12 dB above it, three parallel curves. A fainter fourth curve is the same fan with its blade tone moved into the 2 kHz band, showing the 2 dB increment leaving the 500 Hz band and appearing at 2 kHz. The casing attenuation runs along the bottom from 0 dB at 63 and 125 Hz to 25 dB at 8 kHz, and an inset plots the efficiency correction as a staircase from 16 dB at 40 per cent to 0 dB above 90 per cent](https://docs.jmrp.io/phonometry/media/fan_sound_power.svg)

*The source row, and the two decisions inside it. The efficiency staircase of
the inset is a level shift of the whole spectrum, so a fan selected 25 points
off its best point costs 12 dB before any silencer is priced, more than most
lined duct runs buy back. The blade increment is small (2 dB for this fan type)
but it lands whole in one octave, so what matters is which one. The casing curve
is not a loss on this path at all: it is the power the housing radiates into the
plant room instead of into the duct.*

**Show the code for this figure**

```python
import matplotlib.pyplot as plt

# `fan_sound_power`, `fan_casing_attenuation` and `fan_efficiency_correction`
# are imported above; CFM and IN_WG are the two unit constants.
duty = dict(volume_flow=5000 * CFM, fan_static_pressure_pa=2 * IN_WG,
            fan_type="forward_curved")

# One line for one fan: the band spectrum.
noise_control.fan_sound_power(relative_efficiency_percent=80.0, **duty).plot()
plt.show()

fig, ax = plt.subplots()
for efficiency in (90.0, 80.0, 55.0):
    res = noise_control.fan_sound_power(relative_efficiency_percent=efficiency, **duty)
    ax.semilogx(res.frequencies, res.values, "o-",
                label=f"{efficiency:.0f} % of peak")
casing = noise_control.fan_casing_attenuation()
ax.semilogx(casing.frequencies, casing.values, "s-.", label="Casing")
ax.set_xlabel("Frequency [Hz]"); ax.set_ylabel("Level [dB]")
ax.legend()
plt.show()
```

## 3. What the run takes out

Everything between the fan and the room removes something, and most of it
is free. The models are Long Chapter 14 with the Reynolds (1990)
regressions, and they all return an `HVACSpectrumResult` with `.values`,
`.plot()` and `.report()`. Every model takes the analysis bands as its first
argument, and `None` means the default octave bands from 63 Hz to 8 kHz,
the set the published sheets use; pass your own vector when the sheet runs on
different bands.

**Straight ducts.** An unlined rectangular duct loses energy into the
induced motion of its own walls, so the loss grows with the
perimeter-to-area ratio: a wide shallow duct has floppier side walls than a
square one. `unlined_rectangular_duct_attenuation` fits that below 250 Hz
and holds a flat rate above it; an external fibreglass blanket
(`wrapped=True`) doubles the low-frequency part. A circular duct is far
stiffer in its breathing mode, so it hardly responds at all, and
`unlined_circular_duct_attenuation` is a bare length rate of 0.03 to
0.07 dB/ft. Lining the duct changes the order of magnitude:
`lined_rectangular_duct_attenuation` and
`lined_circular_duct_attenuation` evaluate the Reynolds regressions, valid
for 25 mm to 52 mm linings and clipped at 40 dB per run because flanking
takes over beyond that.

```python
from phonometry import noise_control
import numpy as np

IN, FT = 0.0254, 0.3048
bands = [63.0, 125.0, 250.0, 500.0, 1000.0, 2000.0, 4000.0, 8000.0]

bare = noise_control.unlined_rectangular_duct_attenuation(bands, 36 * IN, 24 * IN, 5 * FT)
lined = noise_control.lined_rectangular_duct_attenuation(bands, 36 * IN, 24 * IN, 5 * FT,
                                                         1 * IN, include_unlined=True)
print(np.round(bare.values, 1))    # [1.1 0.7 0.5 0.2 0.2 0.2 0.2 0.2]
print(np.round(lined.values, 1))   # [ 1.3  1.3  2.5  6.7 12.8 10.6  9.7  9. ]
```

The `include_unlined=True` switch is not cosmetic. The lined-duct
regression was fitted to an *insertion loss*, measured by substituting the
lined section for an unlined one of the same face size, so the side-wall
contribution has been subtracted out of it; Long recommends adding it back
for rectangular ducts, and ignoring it for circular ones where it is
negligible.

**Flexible duct.** The last run of a supply branch is usually flexible
duct, and its published insertion loss is startling: 2 to 3 dB per foot in
the mid bands. `flexible_duct_insertion_loss` interpolates ASHRAE
Table 14.4 over length and log diameter. Part of that number is the duct's
own breakout rather than dissipation, which is exactly why a serpentine run
of flexible duct in a joist space works as an improvised silencer.

**Elbows and splits.** `elbow_insertion_loss` is keyed by $W/\lambda$ and
covers square and round bends, vaned and unvaned, lined and unlined; a
lined square bend is worth 10 to 11 dB where a round one gives 3.
`split_loss` handles a duct division: the power is shared between the
branches in proportion to their areas, plus a reflection when the total
branch area does not match the feeder, and a 25 per cent branch therefore
costs 6 dB.

```python
from phonometry import noise_control

IN = 0.0254
area = 36 * IN * 24 * IN
print(round(noise_control.split_loss(area, [0.25 * area, 0.75 * area], branch=0), 1))  # 6.0
print(noise_control.elbow_insertion_loss(
    [63.0, 125.0, 250.0, 500.0, 1000.0, 2000.0, 4000.0, 8000.0],
    24 * IN, bend_type="round").values)          # [0. 1. 2. 3. 3. 3. 3. 3.]
```

**End reflection.** An open duct end reflects low frequency back up the
run, for free, before any silencer. `end_reflection_loss` offers both
published methods and neither replaces the other: `method="bies"` (the
default) interpolates the ASHRAE table of Bies Table 8.14, and
`method="long"` evaluates Reynolds' closed form

$$
R = 10\log_{10}\!\left[1 + \left(\frac{a\,c}{\pi f d}\right)^{1.88}\right],
$$

with $a = 0.8$ for a flush termination and $a = 1$ for a free one. The two
agree within a decibel or so over the bands both cover. Use
`equivalent_diameter(area)` for a rectangular duct, and do not apply the
correction at all when the duct terminates in a diffuser: the flare smooths
the impedance transition, and a manufacturer's diffuser rating already
contains whatever is left of it.

```python
from phonometry import noise_control
import numpy as np

bands = [63.0, 125.0, 250.0, 500.0, 1000.0, 2000.0]
print(np.round(noise_control.end_reflection_loss(bands, 0.30, method="bies").values, 1))
# [12.  7.  3.  1.  0.  0.]
print(np.round(noise_control.end_reflection_loss(bands, 0.30, method="long").values, 1))
# [12.7  7.7  3.7  1.3  0.4  0.1]
print(round(noise_control.equivalent_diameter(0.36 * 0.24), 3))    # 0.332 m
```

**Silencers and plenums.** A parallel-splitter attenuator reduces, in Bies
§8.10.5, to a set of lined ducts whose liner thickness is half the splitter
thickness, combined by the energy average of Eq. 8.241 so that the leakiest
airway dominates. `plenum_attenuation` is Wells' method for a lined plenum
chamber (Bies Eq. (8.275)), whose reverberant term uses the plenum
[room constant](https://docs.jmrp.io/phonometry/buildings/rooms/room-image-sources/).

```python
from phonometry import noise_control
import numpy as np

IN, FT = 0.0254, 0.3048
sil = noise_control.splitter_silencer_insertion_loss(
    frequencies=None, height_m=24 * IN, length=5 * FT,      # None -> the default octave bands
    airway_widths=[0.10] * 5, splitter_thickness_m=0.10,
)
print(np.round(sil.values, 1))
# [ 5.8  8.6 14.1 28.2 34.5 33.5 18.4 12.1]
print(round(noise_control.plenum_attenuation(0.36, 2.4, 74.0, 0.5), 1))   # 16.1 dB
```

Read that estimate for what it is. The 5 ft unit Long's return path
actually specifies is a low-frequency design worth 16 and 21 dB at 63 and
125 Hz, where this geometry-only model gives 6 and 9. Published dynamic
insertion loss from the manufacturer, measured with the design airflow and
in the design direction to
[ISO 7235](https://docs.jmrp.io/phonometry/devices/noise-control/silencers/#3-how-a-silencer-is-measured-iso-7235),
is what belongs in the sheet; the model is for sizing the airway before there is
a manufacturer.

Element selection is a decision about *shape* (which band does this element buy
me?), so the whole of section 3 belongs on one sheet of paper rather than in six
printed arrays.

![Four octave-band panels from 63 Hz to 8 kHz. Panel a, a 36 by 24 inch run over 5 feet: bare duct falling from 1.1 dB to 0.2, an externally wrapped one starting higher at 2.2 dB and vanishing above 500 Hz, and the same run with 25 mm of lining reaching 12.8 dB at 1 kHz, drawn with and without the side-wall contribution. Panel b, one 24 inch bend in four constructions, the lined square bend reaching 11 dB and the round bend flat at 3, with the W over lambda equals one abscissa marked near 563 Hz. Panel c, the open end of a 300 mm duct by the Bies table and by Long's closed form, agreeing within a decibel from 12 dB at 63 Hz down to zero at 1 kHz. Panel d, the five-airway splitter model rising to 34 dB at 1 kHz, with the two low-frequency values of the unit Long's sheet specifies marked as open diamonds 10 and 12 dB above the model](https://docs.jmrp.io/phonometry/media/duct_attenuation_elements.svg)

*Four families, four different jobs. Lining a run changes the order of
magnitude in the mid bands and does nothing at 63 Hz; a wrapped bare duct does
the opposite, and both are useless above 1 kHz. The elbow tables are keyed on
$W/\lambda$, which is why the same bend behaves alike in ducts of different size
one octave apart. The two end-reflection methods agree within a decibel over the
bands both cover, so the choice between them is not worth an argument. And the
splitter panel is the honest one: the geometry-only regression cannot see the
10 to 12 dB of low-frequency performance a real low-frequency unit is designed
for, which is why that row comes from a datasheet.*

**Show the code for this figure**

```python
import matplotlib.pyplot as plt

# `bands`, IN and FT as above; every model returns an HVACSpectrumResult.

# One line for one element:
noise_control.lined_rectangular_duct_attenuation(bands, 36 * IN, 24 * IN, 5 * FT, 1 * IN,
                                                 include_unlined=True).plot()
plt.show()

# By hand: the bend family of panel (b), keyed on W / lambda.
fig, ax = plt.subplots()
for label, kwargs in (("Square, lined", dict(bend_type="square", lined=True)),
                      ("Square, bare", dict(bend_type="square")),
                      ("Square, vanes", dict(bend_type="square", vanes=True)),
                      ("Round, bare", dict(bend_type="round"))):
    el = noise_control.elbow_insertion_loss(bands, 24 * IN, **kwargs)
    ax.semilogx(el.frequencies, el.values, "o-", label=label)
ax.set_xlabel("Frequency [Hz]"); ax.set_ylabel("Insertion loss [dB]")
ax.legend()
plt.show()
```

One condition travels with the lined-elbow column and is easy to lose: the
tables assume **the duct lining extends at least three duct diameters up- and
downstream of the bend**, so `lined=True` describes a lined run that contains a
bend rather than a lined bend in a bare duct. A bend in bare duct takes the
unlined column; taking the lined one instead overstates the attenuation by the
whole tabulated increment, 10 dB in the mid bands for the square bend above.

## 4. What the system puts back

Attenuation is only half the sheet. Every disturbance of the airflow
generates noise of its own, and past a certain velocity the silencer bought
to remove the fan becomes the loudest thing in the duct.

`silencer_self_noise` is Fry's estimate as Long Eq. 14.31,

$$
L_W = 55\log_{10}\frac{V}{V_0} + 10\log_{10} N
    + 10\log_{10}\frac{H}{H_0} - 45,
$$

spread over the octave bands by the corrections of Table 14.8. The exponent
is the whole message: the fifth-and-a-half power of the airway velocity
means that *doubling the face velocity of a silencer adds about 17 dB*.

```python
from phonometry import noise_control
import numpy as np

IN = 0.0254
slow = noise_control.silencer_self_noise(None, airway_velocity=10.0, n_passages=5,
                                         height_m=24 * IN)
fast = noise_control.silencer_self_noise(None, airway_velocity=20.0, n_passages=5,
                                         height_m=24 * IN)
print(np.round(slow.values, 1))
# [40.8 40.8 38.8 36.8 31.8 26.8 21.8 16.8]
print(round(float(fast.values[0] - slow.values[0]), 1))     # 16.6 dB
```

Straight duct and bends regenerate too, through
`flow_noise_straight_duct` and `flow_noise_bend` (VDI 2081 as Bies
Eqs. (8.251) and (8.254)); the bend model carries the Strouhal-number
transition from the sixth-power inner-corner dipole to the eighth-power
outer-corner quadrupole.

The terminal device is the last one in the path and the one nothing
downstream can fix, because there is no ductwork left after it. Its sound
power is normally manufacturer data measured to ASHRAE Standard 70, and
that is what a real sheet uses. When there is none to hand,
`diffuser_sound_power` is Reynolds's estimate as Long Eqs. 13.27 to 13.33:
an overall level
$L_W = 10\log_{10} S_\mathrm{G} + 30\log_{10}\xi + 60\log_{10} U_\mathrm{G} - 31.3$ from the
face area $S_\mathrm{G}$, the approach velocity $U_\mathrm{G} = Q/S_\mathrm{G}$ and the normalised
pressure-drop coefficient $\xi = 334.9\,\Delta p/(\rho_0 U_\mathrm{G}^2)$ of Eq. 13.28.
Long writes all four in foot-pound units ($S_\mathrm{G}$ in ft², $U_\mathrm{G}$ in ft/s,
$\Delta p$ in inches of water gauge); `diffuser_sound_power` takes and returns
SI and converts internally, which is why the constants here are quoted on
Long's units. The spectrum follows as $L_{W,\text{oct}} = L_W + C_D$ with the
shape function $C_D = -11.82 - 0.15 A - 1.13 A^2$ for a rectangular device
($-5.82$ for a round one), where $A$ is the distance in octaves from the peak
band $f_\mathrm{P} = 48.8\,U_\mathrm{G}$ (Eq. 13.32, $U_\mathrm{G}$ in **ft/s**), counted on Long's band
numbering with band 0 at 32 Hz.

```python
from phonometry import noise_control
import numpy as np

IN, CFM, IN_WG = 0.0254, 0.0004719474432, 249.0

# The supply diffuser of Long's worked sheet: 24 x 24 in, 312 cfm, 0.05 in pd.
print(np.round(noise_control.diffuser_sound_power(None, (24 * IN) ** 2,
                                                  volume_flow=312 * CFM,
                                                  pressure_drop_pa=0.05 * IN_WG).values, 1))
# [ 33.4  32.4  29.1  23.6  15.9   5.9  -6.4 -21. ]
# Long Table 14.9 prints 33/32/29/23/15/4/0/0 for that row.

# The peak band the shape function is counted from, on Long's units:
face_velocity = 312 * CFM / (24 * IN) ** 2 / 0.3048   # 1.3 ft/s
print(round(48.8 * face_velocity, 1))                 # fP = 63.4 Hz
print(round(np.log2(48.8 * face_velocity / 32.0)))    # band 1: the 63 Hz octave
```

The sixth power of velocity in Eq. 13.27 is the design rule: about 18 dB
per doubling of the approach velocity once the pressure drop follows it,
and about 15 dB back for every doubling of face area at the same air
volume. Two screening rules from ASHRAE Chapter 49 come with it.
`air_terminal_velocity_limit` (Table 9) gives the maximum neck velocity for
a design RC, and `air_terminal_damper_correction` (Table 10) gives the
penalty for throttling a balancing damper, which is where a great many
finished installations fail.

```python
from phonometry import noise_control

print(noise_control.air_terminal_velocity_limit(30, opening="supply"))   # 2.2 m/s
print(noise_control.air_terminal_velocity_limit(30, opening="return"))   # 2.5 m/s
print(noise_control.air_terminal_damper_correction(3.0, location="diffuser_neck"))  # 15.0 dB
print(noise_control.air_terminal_damper_correction(3.0, location="supply_duct"))    #  2.0 dB
```

Fifteen decibels in the neck against two decibels 1.5 m back in the duct,
for the same pressure ratio, is the entire design rule: throttle far from
the outlet, or balance the system by sizing the ductwork instead.

Both of this section's models are governed by a slope, and slopes are what
figures are for.

![Two octave-band panels from 63 Hz to 8 kHz. On the left, the self-noise of a five-passage 24 inch silencer at airway velocities of 5, 10, 20 and 40 metres per second: four parallel spectra about 17 decibels apart, running from 24 to 74 decibels at 63 Hz. On the right, the sound power of a 24 by 24 inch diffuser at 312, 624 and 1248 cubic feet per minute, three curves whose peak band moves up one octave for each doubling of the flow, marked by dotted vertical lines at 63, 125 and 250 Hz](https://docs.jmrp.io/phonometry/media/duct_regenerated_noise.svg)

*The two exponents, drawn. The silencer's is the harsher: 55 lg of the airway
velocity puts four parallel curves 16.6 dB apart for every doubling, in every
band at once, which is why a silencer sized for pressure drop rather than for
face velocity becomes the loudest element in the run. The diffuser's peak band
moves as well as its level ($f_\mathrm{P} = 48.8\,U_\mathrm{G}$ in Long's units) so a faster
outlet is not simply louder, it moves its noise into the bands a criterion curve
judges most harshly. The crossing that matters on a real sheet is where one of
these curves meets the attenuated fan level: past it, the element bought to
solve the problem is the problem.*

**Show the code for this figure**

```python
import matplotlib.pyplot as plt

# `silencer_self_noise` and `diffuser_sound_power` as imported above.
bands = [63.0, 125.0, 250.0, 500.0, 1000.0, 2000.0, 4000.0, 8000.0]

# One line for one element:
noise_control.silencer_self_noise(bands, airway_velocity=20.0, n_passages=5,
                                  height_m=24 * IN).plot()
plt.show()

fig, ax = plt.subplots()
for velocity in (5.0, 10.0, 20.0, 40.0):
    res = noise_control.silencer_self_noise(bands, airway_velocity=velocity, n_passages=5,
                                            height_m=24 * IN)
    ax.semilogx(res.frequencies, res.values, "o-", label=f"{velocity:.0f} m/s")
ax.set_xlabel("Frequency [Hz]"); ax.set_ylabel("Regenerated $L_W$ [dB re 1 pW]")
ax.legend()
plt.show()
```

## 4.1 The path that is not in the sheet: breakout

The cascade above accounts only for what stays **inside** the duct. A duct
crossing an occupied space is also a radiating surface, and ASHRAE HVAC
Applications Chapter 49 states that transmission below 250 Hz through duct
breakout is often a major acoustical limitation. The power leaving through the
walls is

$$
L_{W,\text{out}} = L_{W,\text{in}} + 10\log_{10}(S/A) - \mathrm{TL}_\text{out},
$$

with $S = 2L(a+b)$ the radiating surface area of a rectangular duct and
$A = ab$ its cross-section. For this page's own 36 × 24 in trunk over a 3 m run
that geometric term is $10\log_{10}[2 \times 3 \times (0.914 + 0.610)/(0.914
\times 0.610)] = 12.1$ dB, so a wall transmission loss in the teens puts the
radiated power within a few decibels of what is still inside the duct.

The exposure is worst exactly where the sheet still has all its level: at low
frequency, and in the first metres after the fan, before any silencer. A reader
can complete this sheet, pass NC 30, and still fail the room on rumble radiated
by the trunk crossing the ceiling void. The sheet cannot see it, because
breakout leaves the path the sheet is walking. The design rules that follow are
short: get the attenuation in **before** the duct crosses the space, prefer
round or dual-wall duct and a radiused discharge take-off over a
high-aspect-ratio rectangular trunk, and treat lagging as a last resort rather
than a fix, because lagging a rectangular duct is often ineffective. Nothing in
this library models breakout; it is named here so that it is not forgotten.

## 5. From sound power to room level

The last step converts the sound power arriving at the terminal device into
a sound pressure level where somebody is sitting, through the steady-state
room relation
$L_p = L_W + 10\log_{10}\left[Q/(4\pi r^2) + 4/R\right]$. `room_effect`
returns that as a positive attenuation so it drops into the cascade beside
every other loss, with $Q = 2$ by default for a diffuser flush in a
ceiling.

```python
from phonometry import noise_control, room

# The 20 x 20 x 8 ft room of Long's worked sheet, drywall and carpet.
area = 2 * 6.10 * 6.10 + 4 * 6.10 * 2.44        # 134.0 m2
r_const = room.room_constant(area, 0.15)             # 23.6 m2
print(round(float(noise_control.room_effect(1.83, r_const, directivity=2.0)), 1))   # 6.6 dB
```

Long's sheet prints 5 to 7 dB for that room across the bands, so a single
mean absorption of 0.15 lands in the right place; a per-band absorption
gives a per-band room effect, which is what the carpet actually does.

Three of the four inputs deserve naming. The **distance** is the straight line
from the terminal device to the listener's head, and 1.83 m is the 6 ft Long
uses for a seated occupant under a ceiling diffuser; halving it costs 6 dB in
the direct term only, which is why the room effect flattens out once the
reverberant term takes over. The **directivity** $Q$ is 2 for a device flush in
one surface, 4 where a wall meets the ceiling and 8 in a corner, because a
grille in a corner puts its power into an eighth of the sphere. And the formula
is written for **one source**: a room served by four diffusers each carrying a
quarter of the flow is roughly $10\log_{10}4 = 6$ dB louder at a point
equidistant from them than a single-diffuser sheet predicts. The honest
calculation runs one path per outlet and combines them with
`combine_duct_paths`, exactly as the supply and return paths are combined in
section 6. Several outlets over one listener is the commonest reason a passing
sheet fails on site.

### 5.1 The verdict: criterion curve, exceedance and rating

Pass `target=` and `criterion=` to `duct_path` and the result rates itself.
`criterion_curve` samples the NC or RC curve at the analysis bands,
`exceedance` is the band-by-band excess over it, `meets_target` is the
band-by-band verdict a design sheet applies, and `rating` is the full
`NCResult` or `RCResult` derived by the ANSI/ASA S12.2-2019 procedure,
which is a different question and can differ from the tangency verdict. A
spectrum can rate NC-27 and still poke through the NC 30 curve in one band,
because the standard's designation procedure and a sheet's "no band above the
curve" are different questions; `meets_target` answers the second.

That distinction has to be settled before the job is built, because the design
verdict and the commissioning verdict must be the **same quantity**. An
acceptance test specifies: every terminal device serving the space running at
the design flow with the balancing dampers in their final positions (which is
where the 15 dB neck penalty of `air_terminal_damper_correction` actually shows
up, and it is not in any design sheet); microphone positions in the occupied
zone rather than under a diffuser; the system switched off to establish the
background level, with the correction rule agreed in advance; and the rating
procedure applied to the measured octave-band spectrum rather than read off by
tangency. Write into the specification whichever verdict is going to be
measured (band-by-band against the curve, or the S12.2 rating), and not one for
the design and the other for the test.
[Room-noise criteria](https://docs.jmrp.io/phonometry/buildings/rooms/room-noise/) has the rating
procedure itself.

### The same step by VDI 2081

VDI 2081 Blatt 1 closes its own chain with the same relation, Equation (36),
written in the equivalent absorption area $A$ rather than the room constant
$R$. The two are not the same number, $R = A/(1-\bar\alpha)$, so each has its
own argument and neither stands in for the other. The guideline also reads
the directivity off a chart against frequency, so $Q$ may be given one value
per band; a diffuser is more directional the shorter the wavelength.

```python
import numpy as np
from phonometry import noise_control, room

# Room 102 of the guideline's worked example: A = 20 m2, listener 1.5 m away,
# and the directivity its Figure 30 gives for the ceiling diffuser.
q = np.array([2.1, 2.4, 3.0, 4.0, 5.5, 6.7, 7.0, 7.2])
print(np.round(noise_control.room_effect(1.5, absorption_area=20.0, directivity=q), 1))
# [5.6 5.5 5.1 4.7 4.  3.6 3.5 3.4]

# Where the reverberant field takes over, Equation (36a): rH = 0.2 sqrt(A).
print(round(float(room.critical_distance(absorption_area=20.0, directivity=2.0)), 2))  # 0.89
# And the absorption a measured reverberation time implies, Equation (37).
print(round(float(room.sabine_absorption_area(54.4, 0.44, speed_of_sound=339.0)), 1))  # 20.2
```

`sabine_absorption_area` is the familiar Sabine relation read the other way
about, from the time to the area, which is the direction a room survey
actually delivers. Its leading constant follows the speed of sound: $0.161$
at 343 m/s, and the $0.163$ the guideline prints, which is the same
expression at 339 m/s.

Figure 30 itself is not implemented. It is a chart of $Q$ against the product
of frequency and the square root of the outlet area, for two radiation angles
and four positions in the room, and the copy in hand does not resolve it to
better than about a tenth of the quantity it carries. Read it, or take $Q$
from the diffuser's own data, and pass the eight values.

## 6. The worked example: Long's Table 14.9

Long's Chapter 14 closes with a complete sheet: a 5000 cfm forward-curved
fan at 2 in w.g., feeding one room through a supply path (elbow, silencer,
lined duct, a 25 per cent branch split, a second lined duct, flexible duct,
a rectangular diffuser) and a return path (elbow, low-frequency silencer,
lined elbow, plenum, grille), each ending in the room effect of a
20 x 20 x 8 ft carpeted office, combined and checked against NC 30. Every
row below is the one Long prints, including the manufacturer data for the
silencers and the terminal devices, which is what a real sheet uses.

Thirteen rows of a table are hard to hold in the head; the same thirteen rows
as a building are not.

![Sectional schematic of Long's installation. On the left a plant room with the fan, drawn radiating the same sound power into two runs. Above the office ceiling, the supply run left to right as six boxes stamped with the sheet codes 2 to 7: elbow 36 by 24 inches, 3 foot silencer, 5 foot lined duct, 25 per cent branch, 18 by 12 inch lined duct and 6 feet of 12 inch flexible duct, ending in the ceiling diffuser stamped 8. Above it the return run as four boxes stamped 2 to 5, elbow, silencer, lined elbow and plenum, running back to the plant room from the grille stamped 6. On the right the 20 by 20 by 8 foot office with a seated occupant 1.83 metres below the diffuser, marked Q = 2 and NC 30. A legend distinguishes rows that only attenuate, rows that attenuate and regenerate, and rows that are self-noise only](https://docs.jmrp.io/phonometry/media/diagram_duct_path.svg)

*The sheet as a place. The two runs start from the same fan row because a fan
radiates the same power from its intake and its discharge; the codes on the
boxes are the `code=` strings the snippet below stamps on each `DuctElement`,
so a row that fails in the table can be pointed at in the building. The last
line of the legend is the answer this whole sheet arrives at, before any
arithmetic: the return path, not the supply, is what the office hears above
1 kHz.*

```python
import numpy as np
from phonometry import noise_control

bands = [63.0, 125.0, 250.0, 500.0, 1000.0, 2000.0, 4000.0, 8000.0]
fan = [90.0, 86.0, 82.0, 79.0, 77.0, 75.0, 71.0, 61.0]
source = "Fan, centrifugal FC, 5000 cfm, 2 in w.g."

supply = noise_control.duct_path(
    fan, [
        noise_control.DuctElement(label="Elbow, 36 x 24 in, unlined", attenuation=[0, 1, 2, 3, 3, 3, 3, 3],
                                  self_noise=[41, 39, 36, 29, 20, 6, 0, 0], code="2"),
        noise_control.DuctElement(label="Silencer, 3 ft, standard pressure drop",
                                  attenuation=[7, 12, 16, 28, 35, 35, 28, 17],
                                  self_noise=[49, 43, 44, 42, 42, 45, 35, 24], code="3"),
        noise_control.DuctElement(label="Duct, 36 x 24 in, 5 ft, 1 in lining",
                                  attenuation=[2, 2, 3, 7, 15, 12, 11, 9], code="4"),
        noise_control.DuctElement(label="Split, 25 per cent", attenuation=6.0, code="5"),
        noise_control.DuctElement(label="Duct, 18 x 12 in, 6 ft, 1 in lining",
                                  attenuation=[3, 3, 5, 11, 25, 22, 16, 13], code="6"),
        noise_control.DuctElement(label="Flexible duct, 12 in, 6 ft",
                                  attenuation=[14, 14, 16, 15, 17, 22, 16, 13], code="7"),
        noise_control.DuctElement(label="Rectangular diffuser, 312 cfm", attenuation=None,
                                  self_noise=[33, 32, 29, 23, 15, 4, 0, 0], code="8"),
    ],
    frequencies=bands,
    room_effect=[6, 6, 5, 5, 6, 7, 6, 6],
    source_label=source, target=30.0, label="Supply",
)

ret = noise_control.duct_path(
    fan, [
        noise_control.DuctElement(label="Elbow, 36 x 24 in, unlined", attenuation=[0, 1, 2, 3, 3, 3, 3, 3],
                                  self_noise=[43, 42, 39, 33, 24, 12, 0, 0], code="2"),
        noise_control.DuctElement(label="Silencer, 5 ft, low-frequency type",
                                  attenuation=[16, 21, 35, 41, 41, 28, 21, 15],
                                  self_noise=[51, 49, 53, 56, 56, 59, 60, 53], code="3"),
        noise_control.DuctElement(label="Elbow, 36 x 24 in, lined, 1 in", attenuation=[1, 2, 3, 4, 5, 6, 8, 10],
                                  self_noise=[39, 38, 34, 28, 18, 4, 0, 0], code="4"),
        noise_control.DuctElement(label="Plenum, 800 sq ft, 50 per cent lined",
                                  attenuation=[12, 13, 19, 20, 20, 20, 21, 21], code="5"),
        noise_control.DuctElement(label="Rectangular grille, 24 x 24 in, 563 cfm", attenuation=None,
                                  self_noise=[30, 29, 26, 20, 12, 1, 0, 0], code="6"),
    ],
    frequencies=bands,
    room_effect=[9, 8, 6, 8, 8, 8, 9, 10],
    source_label=source, target=30.0, label="Return",
)

total = noise_control.combine_duct_paths([supply, ret], label="Supply + return")
print(np.round(supply.received_level, 0))    # Long: 52 42 30 18  9 -2 -2 -1
print(np.round(ret.received_level, 0))       # Long: 52 41 27 25 23 25 22 12
print(np.round(total.received_level, 0))     # Long: 55 45 32 26 23 25 22 12
print(total.meets_target)                    # True
print(round(float(total.rating.rating), 1))  # 26.7, governed at 63 Hz
```

Every printed row comes back within the sheet's own 1 dB rounding: the
supply is 1 dB low at 4 kHz, the return 1 dB low at 500 Hz, and the
combination 1 dB low at 500 Hz and 1 dB high at 8 kHz, everything else
exact. The room lands at NC 27, comfortably inside its NC 30 target, and
the 63 Hz band is what governs the rating, which is the usual outcome of a
duct-noise design and the reason low-frequency silencer performance is
worth paying for.

![Octave-band levels delivered into the room by the supply and return paths of Long's worked HVAC sheet, together with their energy sum and the NC 30 criterion curve: both paths start near 52 dB at 63 Hz, the supply falls away steeply above 500 Hz to below 0 dB while the return stays flat near 23 to 25 dB, and the combined received spectrum runs a few decibels under the NC 30 curve in every band](https://docs.jmrp.io/phonometry/media/duct_path_cascade.svg)

*The two paths and their sum against NC 30. The supply, with its silencer,
two lined runs, a branch split and six feet of flexible duct, has nothing
left above 1 kHz; the return, with a plenum but a silencer whose own
self-noise floors it near 25 dB, is what the room actually hears in the mid
and high bands. Adding low-frequency attenuation to the supply would change
nothing at all: the return already sets the answer everywhere except at
63 Hz, and that is the row to argue about.*

**Show the code for this figure**

```python
import matplotlib.pyplot as plt
from phonometry import noise_control

# `supply` and `ret` are the two DuctPathResult objects built above.

# One line for one path: the cascade of the supply run against NC 30.
supply.plot()
plt.show()

# The concept figure: both paths, their energy sum and the criterion.
total = noise_control.combine_duct_paths([supply, ret], label="Supply + return")
total.plot()
plt.gca().set_ylim(-6.0, 62.0)
plt.show()
```

`DuctPathResult` also prints and files itself. `.table()` returns the sheet
row by row with the worksheet sign convention, and `.report()` renders a
one-page PDF in the layout of the published procedures (AHRI Standard 885
Table 8; Long Table 14.9): the element table, the cascade chart against the
criterion curve, the boxed room-criterion rating and the verdict.

```python
for row in total.table():
    print(f"{row['kind']:<12} {row['label'][:26]:<26} "
          f"{[round(float(v)) for v in row['values']]}")

supply.report("duct-path-supply.pdf")    # the fiche previewed below
total.report("duct-path.pdf")            # the combination; needs phonometry[report]
```

The example fiche is the **supply half** of the sheet, regenerated with
`make reports` and kept rendered in the repository. Click the preview to open
the PDF:

**Report:** [Duct-borne noise path calculation example report (PDF)](https://docs.jmrp.io/phonometry/media/duct_path_example.pdf): Duct-borne noise path calculation example report: a metadata header with the client, the noise source, the test environment and the date, the octave-band path table listing the fan sound power, each element attenuation as a negative level change, the self-noise rows of the elbow, the silencer and the diffuser, the room effect, the received level and the NC 30 curve, and beneath it the boxed room criterion NC-22.6 at 125 Hz with the verdict that no band exceeds NC 30 beside the cascade chart of every element against the criterion curve.

Duct-borne noise path fiche (DuctPathResult.report): the supply path of the worked sheet element by element, rated NC-22.6, governed at 125 Hz, and passing its NC 30 target in every band.

## 7. What reproduces, and what does not

Long's Table 14.9 was produced by a commercial computer program, not by
hand from the tables printed alongside it, and being honest about that is
more useful than pretending otherwise. The *arithmetic* of the sheet is
reproduced exactly, as section 6 shows. Several of its **element rows**,
however, do not follow from the book's own printed data, and the functions
in this library implement the printed equations and tables. Verified band
by band:

- **The fan row does not come from Eq. 13.1.** The sheet prints
  90/86/82/79/77/75/71/61 dB. Eq. 13.1 with the Table 13.5 forward-curved
  constants at 5000 cfm and 2 in w.g. gives 99/99/89/84/82/77/72/67 dB, and
  the printed spectrum is not a level shift of the tabulated one, so it
  comes from other data (a manufacturer's, most likely).
- **The flexible-duct row is not Table 14.4.** The sheet prints
  14/14/16/15/17/22/16/13 dB for 12 in by 6 ft;
  `flexible_duct_insertion_loss` reads 3/5/10/15/17/16/9 dB out of the
  table for that duct.
- **The lined rectangular ducts agree in the mid and high bands.** For the
  18 x 12 in, 6 ft, 1 in run the library returns 11/25/22/16/13 dB from
  500 Hz up, exactly the printed row, and is 1 to 2 dB high below it
  (5/4/6 against 3/3/5). For the 36 x 24 in, 5 ft run it matches at 250,
  500 and 8 kHz and is 1 to 2 dB low elsewhere.
- **The split and the unlined elbow reproduce exactly.** `split_loss` gives
  the 25 per cent branch as 6.0 dB against the printed -6 dB, and
  `elbow_insertion_loss` gives 0/1/2/3/3/3/3/3 dB against the printed row
  when the elbow is read as round (Table 14.7) at $w = 24$ in.
- **The supply diffuser row reproduces too.** `diffuser_sound_power` on a
  24 x 24 in rectangular device at 312 cfm and 0.05 in pd returns
  33.4/32.4/29.1/23.6/15.9/5.9 dB against the printed
  33/32/29/23/15/4, inside the sheet's own rounding in the five bands that
  carry the level. The return grille row (30/29/26/20/12/1) does not follow
  from the same equations at its 563 cfm, so it is manufacturer data.
- **The NC 30 row differs by 1 dB at 1 kHz.** Long prints
  57/48/41/35/31/29/28/27; the library's `nc_curve(30)` returns
  57/48/41/35/**32**/29/28/27, the values of ANSI/ASA S12.2-2019 Table 1.
  Long is using the original Beranek 1957 curve. The difference does not
  change the verdict here, but it is worth knowing which NC you are
  quoting.

![Six octave-band panels comparing Long's printed Table 14.9 rows, drawn as filled markers, with the library's computed values drawn as a line, the difference shaded between them. The fan panel disagrees by up to 13 dB and the printed spectrum is not a level shift of the computed one; the flexible-duct panel disagrees by up to 11 dB with the printed row much higher at low frequency; the two lined ducts agree within 2 dB; the supply diffuser agrees within 1 dB in every band that carries level; and the NC 30 curve differs by 1 dB in the 1 kHz band alone](https://docs.jmrp.io/phonometry/media/duct_sheet_verification.svg)

*Where the sheet is a calculation and where it is a datasheet, seen at a glance.
The two lined runs and the diffuser sit inside the sheet's own rounding, so those
rows really are the printed equations. The fan and the flexible duct do not, and
the shape of the disagreement says why: the fan's printed spectrum is not a level
shift of the tabulated one, so it comes from other data, and the flexible-duct
row is far above ASHRAE Table 14.4 at low frequency, which is where its breakout
lives. The last panel is the one to remember when quoting a rating: Long's NC 30
is the original Beranek 1957 curve and differs from ANSI/ASA S12.2-2019 Table 1
by 1 dB, in the 1 kHz band alone.*

**Show the code for this figure**

```python
import matplotlib.pyplot as plt

# Every model is imported above; nc_curve returns the 10-band curve aligned
# with room.NC_OCTAVE_BANDS_HZ, which starts at 16 Hz, so the sheet's eight bands are
# nc_curve(30)[2:].
printed = [3, 3, 5, 11, 25, 22, 16, 13]      # Long's 18 x 12 in, 6 ft row
computed = noise_control.lined_rectangular_duct_attenuation(
    bands, 18 * IN, 12 * IN, 6 * FT, 1 * IN, include_unlined=True).values

fig, ax = plt.subplots()
ax.fill_between(bands, printed, computed, alpha=0.25)
ax.semilogx(bands, printed, "o", label="Long's printed row")
ax.semilogx(bands, computed, "-", label="the library")
ax.set_xlabel("Frequency [Hz]"); ax.set_ylabel("Level [dB]")
ax.legend()
plt.show()
```

None of that is a defect of the sheet. It is what a real duct-borne
calculation looks like: the elements a manufacturer publishes (fans,
silencers, diffusers, grilles) come from test data, and the elements nobody
publishes (duct runs, elbows, splits, end reflections, the room) come from
the tables. `DuctElement` takes both without caring which is which, which
is the point.

## 8. The plane-wave limit

Every element model above, and every four-pole silencer in
[Silencers](https://docs.jmrp.io/phonometry/devices/noise-control/silencers/), is one-dimensional. It assumes
a single sound pressure describes the whole duct cross section, which is
true only below the frequency at which the first higher-order acoustic mode
cuts on. Above it several modes propagate at once, each with its own axial
wavenumber, and a plane-wave prediction quietly stops being right.

`noise_control.duct_modes` implements the cut-on analysis of Norton &
Karczub, *Fundamentals of Noise and Vibration Analysis for Engineers*
(2nd ed.), section 7.3: circular ducts by Eq. 7.6 with the
$\pi\alpha_{pq}$ eigenvalues of Table 7.1 that solve
$J'_p(\kappa_{pq} a_\mathrm{i}) = 0$, rectangular ducts by Eq. 7.10,

$$
(f_\text{co})_{pq} = \frac{c}{2}\sqrt{(p/a)^2 + (q/b)^2}\,\sqrt{1 - M^2},
$$

and the mean-flow correction of Eqs. 7.8 and 7.9, in which a uniform axial
flow of Mach number $M$ lowers every cut-on frequency by $\sqrt{1 - M^2}$
and moves the cut-on itself from $k_x = 0$ to
$k_x = -M\kappa_{pq}/\sqrt{1 - M^2}$.

```python
import numpy as np
from phonometry import noise_control

# Norton problem 7.2: a 0.65 x 0.4 m air-conditioning duct at 15 m/s.
modes = noise_control.rectangular_duct_cut_on(0.65, 0.40, flow_velocity=15.0, n_modes=6)
print(modes.modes[:3])                          # ((1, 0), (0, 1), (1, 1))
print(np.round(modes.cut_on[:3], 1))            # [263.6 428.3 502.9] Hz
print(np.round(modes.cut_on_no_flow[:3], 1))    # [263.8 428.8 503.4] Hz
print(round(modes.plane_wave_limit, 1))         # 263.6 Hz

IN = 0.0254
print(round(noise_control.plane_wave_limit(width_m=36 * IN, height_m=24 * IN), 1))   # 187.6 Hz
print(round(noise_control.plane_wave_limit(diameter_m=12 * IN), 1))                # 659.5 Hz
```

Those ventilation numbers are blunt: in that duct plane waves are the whole
story only up to the 250 Hz octave, and a 36 x 24 in supply trunk gives up at
188 Hz.

What a number cannot say is that cut-on is a **switch**, not a slope. The clip
below simulates the 0.65 m dimension of that same duct as a 2D slice, rigid
walls, and drives it twice with one off-axis source that excites the plane
mode and the first transverse mode together: once at 180 Hz, below the
263.8 Hz the snippet just printed, and once at 400 Hz above it. Nothing else
differs. Below cut-on the transverse mode is *evanescent*: it decays as
$\mathrm{e}^{-\alpha x}$ with $1/\alpha = 0.28$ m, which puts it 20 dB down 0.65 m from
the source (one duct width), so the lumpy near field dies there and every
wavefront after it crosses the section flat. Above cut-on it propagates, the two modes run at different
axial wavenumbers ($k_x = 5.51$ against $k = 7.33$ rad/m at 400 Hz) and beat
along the duct every 3.45 m, and the profile across the section never settles.
The panel beside each strip draws that profile at a fixed station 5 m
downstream against its own section average, which is the single pressure the
one-dimensional models above carry, so "a single sound pressure describes the
whole cross section" is either true on screen or visibly false.

**Video:** [A rigid duct driven below and above its first cut-on frequency](https://docs.jmrp.io/phonometry/media/anim_fdtd_duct_cut_on.webm): The same 0.65 metre rigid duct is driven at 180 Hz and at 400 Hz by one off-axis source. In the 180 Hz strip the lumpy field near the source flattens within half a duct width and every later wavefront crosses the section as a straight line, and the profile panel beside it collapses onto its own average. In the 400 Hz strip the field stays lumpy the whole length of the duct, the two propagating modes beat along it, and the profile panel swings far away from the average on every cycle.

That is why a transfer-matrix silencer model has a **ceiling frequency**
rather than an accuracy budget: below cut-on the second mode is not small,
it is absent, and above it the model is describing one of several waves that
are all there. The
[Silencers](https://docs.jmrp.io/phonometry/devices/noise-control/silencers/) four-pole algebra
inherits the same limit.

At 15 m/s the flow correction is invisible: $M = 0.044$ gives
$\sqrt{1 - M^2} = 0.999$, which moves the first cut-on by 0.2 Hz. It earns
its place in high-speed pipework instead. Norton's problem 7.1 is that
case, a 254 mm line carrying steam ($c = 405$ m/s) at 200 m/s, $M = 0.494$,
and there the two ladders separate by more than a hundred hertz at every
rung.

```python
import numpy as np
from phonometry import noise_control

# Norton problem 7.1: a 254 mm circular duct carrying steam at 200 m/s.
steam = noise_control.circular_duct_cut_on(0.254, flow_velocity=200.0,
                                           speed_of_sound=405.0, n_modes=6)
print(steam.modes)
# ((1, 0), (2, 0), (0, 1), (3, 0), (4, 0), (1, 1))
print(np.round(steam.cut_on_no_flow, 1))
# [ 934.5 1550.1 1944.7 2132.3 2698.9 2705.9]
print(np.round(steam.cut_on, 1))
# [ 812.6 1347.9 1691.1 1854.1 2346.8 2352.9]
print(np.round(steam.axial_wavenumber, 2))
# [ -8.23 -13.66 -17.13 -18.79 -23.78 -23.84]
```

![Cut-on frequency of the first six higher-order acoustic modes of a 254 mm circular steam line, plotted against the mode order (1,0), (2,0), (0,1), (3,0), (4,0) and (1,1): the still-air ladder climbs from 935 Hz to 2706 Hz as a grey dashed line while the 200 m/s ladder runs 12 per cent below it from 813 Hz to 2353 Hz, and the band below the first cut-on is shaded as the plane-wave-only region](https://docs.jmrp.io/phonometry/media/duct_mode_cut_on.svg)

*Norton's problem 7.1, the case where the mean flow is worth drawing: half
the speed of sound in the pipe pulls every cut-on down by
$\sqrt{1 - M^2} = 0.870$, so the first higher-order mode appears at 813 Hz
instead of 935 Hz and the plane-wave band, shaded, is 13 per cent narrower
than the still-air calculation would promise. The axial wavenumber at cut-on
is negative in every rung: with flow, the mode is already travelling upstream
at the frequency at which it appears.*

**Show the code for this figure**

```python
import matplotlib.pyplot as plt
from phonometry import noise_control

# One line for one duct: the cut-on ladder with the plane-wave band shaded.
steam = noise_control.circular_duct_cut_on(0.254, flow_velocity=200.0,
                                           speed_of_sound=405.0, n_modes=6)
steam.plot()
plt.show()

# The ventilation duct of problem 7.2, where the flow shift is negligible.
noise_control.rectangular_duct_cut_on(0.65, 0.40, flow_velocity=15.0, n_modes=6).plot()
plt.show()
```

Two results carry this limit for you. Every `ReactiveSilencerResult`
reports the first cut-on of its widest cross section as
`plane_wave_limit`, and `duct_path` accepts a `section=` description of the
duct it is walking. Both raise a `PlaneWaveWarning` when the analysis grid
runs past that frequency: the numbers are still returned, and above cut-on
they describe the plane-wave mode alone, which a measurement will not.

```python
import warnings
from phonometry import noise_control

IN = 0.0254
bands = [63.0, 125.0, 250.0, 500.0, 1000.0, 2000.0, 4000.0, 8000.0]

with warnings.catch_warnings(record=True) as caught:
    warnings.simplefilter("always")
    noise_control.duct_path([90.0] * 8, [noise_control.DuctElement(label="Straight run", attenuation=3.0)], frequencies=bands,
                            section={"width_m": 36 * IN, "height_m": 24 * IN},
                            flow_velocity=6.0, label="Supply")
print(caught[0].category is noise_control.PlaneWaveWarning)
print(str(caught[0].message))
# Supply: 6 of 8 frequencies are above the first duct cut-on frequency
# (188 Hz), where higher-order modes propagate and the plane-wave result
# describes the plane-wave mode only.
```

Six of the eight octave bands of a standard duct-noise sheet sit above the
cut-on of a 36 x 24 in duct. That is not a reason to distrust the sheet:
the ASHRAE element models it is built from are *empirical*, fitted to
measurements of real ducts in which those modes were present, so they carry
the multimode behaviour inside their regression constants. The warning is
aimed at the analytical methods, the four-pole silencer algebra above all,
where the plane-wave assumption is structural rather than statistical, and
where the peaks and troughs of a computed transmission loss simply do not
survive past cut-on.

## What this guide covers

- **Covered.** The fan source model (`fan_sound_power`, `fan_efficiency_correction`,
  `blade_passing_frequency`, `fan_casing_attenuation`; Long Eq. 13.1 with Tables
  13.5-13.8); the run attenuations (lined and unlined rectangular and circular
  ducts, `flexible_duct_insertion_loss`, `elbow_insertion_loss`, `split_loss`,
  `end_reflection_loss` in both published methods,
  `splitter_silencer_insertion_loss`, `plenum_attenuation`); the regenerated
  noise (`silencer_self_noise`, `flow_noise_straight_duct`, `flow_noise_bend`,
  `diffuser_sound_power` and the ASHRAE Chapter 49 screening rules
  `air_terminal_velocity_limit` and `air_terminal_damper_correction`);
  `room_effect` and the criterion machinery; the cascade itself (`duct_path`,
  `combine_duct_paths`, `.table()`, `.report()`); and the `duct_modes` cut-on
  analysis.

- **Not covered.** No fan or terminal-device data is predicted where the manufacturer publishes
  it, and this sheet expects the published figure. There is no dissipative-liner
  model: the splitter estimate is a lined-duct regression and the plenum attenuation
  is Wells' closed form. Above the first cut-on the element models stand on their
  empirical fit alone. And **four noise paths of a real installation are outside
  this calculation entirely**, which is why a passing sheet and a failing room
  are usually one of them:

  - **Duct breakout**, treated in section 4.1: a large low-frequency-carrying
    trunk radiating through its own walls into a space it merely crosses,
    frequently the governing path near a plant room.
  - **Duct break-in and crosstalk**: two rooms served by the same run hearing each
    other through it, regardless of the partition between them.
  - **Structure-borne transmission of fan vibration** through hangers, plinths
    and the slab, which no airborne sheet can see and which is treated with
    isolators, and whose dynamic transfer stiffness is measured in
    [Transfer stiffness (ISO 10846)](https://docs.jmrp.io/phonometry/vibration/structural/transfer-stiffness/).
  - **The octave-band resolution itself**, which averages a blade-passing tone
    into its band and so understates how a pure tone will be judged.

## See also

- [Silencers](https://docs.jmrp.io/phonometry/devices/noise-control/silencers/): the reactive four-pole
  elements (expansion chambers, Helmholtz, quarter-wave and extended-tube
  resonators) whose validity ends at the same cut-on frequency.
- [Industrial noise control](https://docs.jmrp.io/phonometry/devices/noise-control/noise-control/): the rest
  of the installation, the individual HVAC duct methods and
  machine-enclosure insertion loss.
- [Room-noise criteria (NC / RC Mark II)](https://docs.jmrp.io/phonometry/buildings/rooms/room-noise/):
  the ANSI/ASA S12.2-2019 families the received spectrum is judged against.
- [Steady-state room field](https://docs.jmrp.io/phonometry/buildings/rooms/room-image-sources/): the
  room constant behind the room effect and the plenum reverberant term.
- API reference:
  [`noise_control`: duct path](https://docs.jmrp.io/phonometry/reference/api/noise_control/duct-path/),
  [`noise_control.duct_modes`](https://docs.jmrp.io/phonometry/reference/api/noise_control/duct-modes/),
  [`noise_control.hvac`](https://docs.jmrp.io/phonometry/reference/api/noise_control/hvac/).

## References

- **Air-Conditioning, Heating and Refrigeration Institute. (n.d.).** *Procedure for estimating occupied space sound levels in the application of air terminals and air outlets* (AHRI Standard 885).
  The industry row structure of the duct-borne calculation sheet (Table 8) that DuctPathResult.table() and .report() follow.
- **American Society of Heating, Refrigerating and Air-Conditioning Engineers. (2019).** *ASHRAE handbook: HVAC applications (SI edition), Chapter 49, Noise and vibration control*. ASHRAE.
  The air terminal velocity limits (Table 9), the volume-damper corrections (Table 10) and the guidance that fan sound power is best taken from manufacturer test data.
- **Bies, D. A., Hansen, C. H., & Howard, C. Q. (2017).** *Engineering noise control* (5th ed.). CRC Press. https://doi.org/10.1201/9781351228152
  The end-reflection table (§8.13), the elbow insertion loss (§8.11), the reduction of a splitter muffler to lined ducts (§8.10.5, Eq. 8.241), the plenum chamber (§8.17) and the flow-generated noise of ducts and bends (§8.15).
- **Long, M. (2014).** *Architectural acoustics* (2nd ed.). Academic Press. https://doi.org/10.1016/C2012-0-03257-5
  The fan sound-power model (Ch. 13, Eq. 13.1 with Tables 13.5-13.8) and the diffuser self-noise model (Ch. 13, Eqs. 13.27-13.33), the duct attenuation, flexible duct, split loss, end reflection, silencer self-noise and room effect of Ch. 14, and the worked duct-borne sheet of Table 14.9 this guide is built around.
- **Norton, M. P., & Karczub, D. G. (2003).** *Fundamentals of noise and vibration analysis for engineers* (2nd ed.). Cambridge University Press. https://doi.org/10.1017/CBO9781139163927
  The higher-order acoustic modes of ducts, the cut-on frequencies of circular and rectangular cross sections and the mean-flow correction (§7.3, Eqs. 7.6-7.10).
- **Verein Deutscher Ingenieure. (2001).** *Geräuscherzeugung und Lärmminderung in Raumlufttechnischen Anlagen* (VDI 2081 Blatt 1:2001-07).
  Section 6.7.3: the level an air outlet produces in a room (Equation (36)), the equivalent absorption area a reverberation time implies (Equation (37)) and the reverberation radius rH = 0.2 sqrt(A). Element 20 of the worked example in its Part 2:2005-05 is the case checked here.
