Environmental noise prediction has become very good. The calculation methods are standardised, the software implementations are mature, and two competent practitioners given the same inputs will produce closely comparable results.

That last clause carries the whole difficulty. The model is a faithful reporting instrument. It reports the consequences of what it was given, with considerable precision, and it does so whether or not what it was given resembles the project.

Two things get treated as one here, and they are not the same. Modelling uncertainty is the irreducible spread associated with a calculation method applied within its stated limits, and it can be quantified and carried through an assessment. Input error is a wrong assumption faithfully propagated, and it does not appear as uncertainty at all. It appears as a confident number.

The five gaps below are input errors, not uncertainty. Each one is recoverable at design stage and expensive after.

One, source data that is not the quantity it appears to be 

More assessments are compromised here than anywhere else, and the failure has several distinct forms.

The most consequential is the substitution of a sound pressure level for a sound power level. Sound power is a property of the source, independent of distance and surroundings. Sound pressure depends on distance, directivity, reflecting surfaces and the space the measurement was taken in. A figure quoted as “75 dB(A) at 1 m” is a sound pressure level under conditions that are frequently not stated, and converting it to sound power requires knowing what those conditions were.

The second is the grade of the data. Sound power determined under ISO 3744 by an engineering grade method and sound power determined under ISO 3746 by a survey method are not equivalent inputs. ISO 3746 is explicitly classed as accuracy grade 3, the lowest of the three grades, and it says so in its own introduction. Both are legitimate. They should not be mixed within a plant schedule without the assessment saying so.

The third is spectral content. A source supplied as a single A-weighted figure cannot be propagated correctly, because atmospheric absorption, ground attenuation and barrier performance are all strongly frequency dependent. ISO 9613-2 performs its calculation in octave bands from 63 Hz to 8 kHz for exactly this reason. Where octave band data is unavailable and a generic spectrum is substituted, the substitution has usually been made at the point in the spectrum where large mechanical plant differs most from the generic case.

The fourth is the operating point the data represents. A fan or engine sound power level applies at a stated speed and load. Part load is not proportionally quieter, and on variable speed plant the relationship between duty and radiated sound power is neither linear nor always monotonic across bands.

Two, operating modes that do not match how the plant will actually run 

A model runs the scenarios it is given. Which scenarios it is given is a project decision, not an acoustic one, and it is frequently made with incomplete information about the operating regime.

The recurring gaps are the same across sectors. Redundant plant that is assessed as standby and in practice runs simultaneously during changeover, testing or summer peak conditions. Plant that is assessed at design duty and operates at a higher duty because the design allowed margin. Maintenance and testing regimes, particularly for standby generation, that were never presented as an operating case at all.

Duty cycle also interacts with the assessment metric in a way that is easy to get wrong. An event that fills a short assessment interval receives no reduction for being short. The same event assessed over a long averaging period does. Whether duration helps depends entirely on which metric the applicable framework specifies, and both metrics frequently apply at once.

Three, receivers described too coarsely to be meaningful 

A receiver in a model is a coordinate, a height and a criterion. Each of the three is capable of being wrong in a way that changes the outcome.

Location is the obvious one, and the recurring version is assessing at the site boundary when the criterion applies at a dwelling, or the reverse.

Height is less obvious and often more consequential. Ground attenuation and barrier performance both depend on the geometry of the propagation path, and a first floor bedroom window frequently sees less screening than a ground floor position at the same address. Where a barrier or building is providing much of the predicted attenuation, the receiver height is doing a lot of work in the result.

The measurement convention is a third source of divergence. Prediction and measurement conventions differ on whether a receiver level is expressed as a free field level or includes the reflection from the building facade behind the microphone, a difference of the order of three decibels. An assessment that predicts on one convention and is later verified on the other will appear to have a discrepancy that is not real.

Finally, receivers change. Approved but unbuilt dwellings, rezoned land and future subdivisions are foreseeable receivers, and an assessment based only on what is currently constructed can be overtaken by a planning decision that was already in the public record when the model was built.

Four, terrain, ground and screening treated as simple objects 

The propagation part of the model tends to receive less scrutiny than the source part, because it looks like geometry rather than data. It contains several assumptions that materially move the result.

The ground factor determines how much attenuation the intervening ground provides, from acoustically hard surfaces at one extreme to porous ground at the other. Sites frequently contain both, and they change during construction. A site modelled with soft ground and delivered with hardstand, pavement and detention basins is a different acoustic surface.

Screening assumptions are the most consequential. A barrier in a model provides its calculated diffraction attenuation on the assumption that transmission through it is negligible and there are no gaps. A real barrier with a modest open area, whether from panel joints, drainage gaps, a gate or a service penetration, has its performance capped at a level well below the diffraction calculation. This is the same composite mechanism that limits enclosure performance, described in our article on choosing the right acoustic metric.

Directivity is routinely simplified. A radiator discharge, a stack outlet, a louvred aperture and an unattenuated opening are all strongly directional sources. Modelling them as omnidirectional points distributes their energy in a way that under predicts the receiver on the axis and over predicts the receivers off it. Buildings modelled as simple blocks are a related simplification, providing screening in the model and also providing reflections that are sensitive to how the surfaces were described.

Five, a scope boundary drawn around today’s project 

The final gap is not about acoustics at all. It is about what was included.

Where a site is staged, an assessment covering only the current stage produces a valid answer to a narrow question and an invalid basis for the equipment selection, because the noise budget that stage is consuming is the same budget the next stage will need. Where existing industrial noise is already present, an assessment that models only the new plant is answering a different question from the one the amenity criterion asks.

Both of these are cumulative problems, and they behave logarithmically. Several sources that individually satisfy a criterion combine to exceed it, and the arithmetic is covered in our article on cumulative plant noise.

Meteorology, which sits underneath all five 

Propagation is meteorologically conditioned, and it is easy to misread what a standard method already assumes.

ISO 9613-2 (2024) provides an engineering method that addresses downwind propagation, or propagation under a well developed moderate ground based temperature inversion of the kind that occurs on clear calm nights. That condition is built into the method. It is not a neutral prediction to which adverse propagation conditions are added afterwards, and it is not automatically the worst case for every site, since the standard notes that inversion conditions over extended water surfaces are not covered and can produce higher levels than predicted. Its accuracy and usage constraints are stated in the standard rather than left to the user.

Australian frameworks add their own requirements on top. The NSW Noise Policy for Industry requires assessment under standard conditions and under the meteorological conditions the policy identifies as increasing received noise levels, with inversion conditions identified using Pasquill-Gifford stability categories, and requires that the model used be capable of considering them.

An assessment that does not state which meteorological condition each predicted level represents has produced numbers that cannot be compared with a criterion.

What changes when the inputs are wrong 

A model output is not a report. It is an instruction to the rest of the project.

Predicted receiver levels set the mitigation scope. Mitigation scope sets the equipment selection. Equipment selection sets plantroom space, facade openings, structural loads and duct routes. Those set the layout, the layout sets what the tender package contains, and the tender package sets what the project is contractually able to build.

An input error at the top of that chain does not stay at the top. It arrives as a plantroom that cannot accommodate the attenuator length the corrected model requires, on a project where the concrete is poured.

Supplier data belongs in the same discussion 

Model inputs are not only source data. The performance of every mitigation element in the model is also an input, and it deserves the same scrutiny.

Three questions belong with any supplied performance figure. Is it octave band data or a single number rating, given that the model needs bands. Does it apply to the configuration and the operating condition being modelled, particularly for airstream components whose insertion loss is stated at a specific flow. And what class of evidence is behind it, whether laboratory measurement under an accredited scope, or calculation aligned to a test method, both of which are legitimate and are not the same.

AcousTech publishes its Sonic System acoustic modular panel and Sonic Access acoustic door ratings as laboratory results, and its Sonic Series acoustic louvre performance assessed to AS 1191-2002 and rated to AS/NZS ISO 717.1, with full performance data available on request. A model that treats those two classes of input as equally certain has introduced an assumption its author did not intend to make.

None of this is an argument that acoustic modelling is unreliable. The calculation methods are well characterised, their limitations are published, and within their stated scope they perform well.

The argument is narrower and harder to act on. A model is an amplifier for the quality of the project information available when it was built, and the sophistication of the method has no bearing on the quality of what goes into it. Refining a propagation calculation while the source spectra are catalogue estimates, the operating case is nominal and the second stage of the site is outside the model is precision applied to the wrong part of the problem.

The most useful thing an assessment can contain is a clear statement of what was assumed, which assumptions the result is sensitive to, and what would have to be confirmed to reduce that sensitivity. That turns the model from a number into an engineering document, and it gives the project team something they can act on while acting is still inexpensive.