PaveKeep Sign in
Knowledge base

How road deterioration is modelled, and what a model cannot tell you

Projecting a network ten years forward is arithmetic, not prophecy. The useful question is not how clever the model is — it is which of its inputs your council actually holds.

Every ten-year capital plan contains a projection. A council states what the network will look like under one spending level and under another, and the difference between them is the argument for the budget. Those projections come out of a deterioration model, and most of the time nobody in the room has seen what is inside it.

That is a problem when the projection is challenged, because the answer "the software calculated it" concedes the point.

What a deterioration model is doing

At its simplest it answers one question per segment: given this surface, in this condition, carrying this traffic, how much worse will it be in a year?

Austroads sets out what a proper treatment-design investigation draws on (Guide to Pavement Technology Part 5, 2025): the original pavement design, construction details, maintenance and rehabilitation records, climatic conditions, and the effect of traffic on past performance. Those are the inputs that explain why two visually identical roads behave differently.

Most small councils hold two of the five. That is not a failure — it is the condition the method has to work in, and a model that quietly assumes the other three is making things up on your behalf.

Three honest limits

1. A model projects; it does not predict. It applies a published rate to a measured condition and reports where that arithmetic lands. Everything it produces is conditional on the rate being right for your roads, which is a question about your roads and not about the software.

2. Heavy vehicles matter more than vehicle counts. Pavement damage scales roughly with the fourth power of axle load, so a route with a modest traffic count and a high heavy-vehicle share can deteriorate faster than a busier street. A model that takes only AADT is taking the wrong number. One that takes a heavy-vehicle share should say so visibly, because it is a large effect driven by a figure many councils estimate rather than measure.

3. The things a model leaves out do not stop existing. Climate — rainfall, temperature, frost — is a genuine driver of deterioration and is absent from most local-government models, including ones that look sophisticated. So is subgrade. A model that omits them is still useful; a model that omits them silently lets a reader believe the projection accounts for things it has never seen.

The test that actually matters

Not "how accurate is the model", which cannot be answered in advance, but:

Can somebody who disagrees with the projection see exactly what produced it?

That means the rates are published rather than embedded, the thresholds are visible, the inputs per segment are inspectable, and where a number is modelled rather than measured it is labelled as modelled wherever it appears. A council officer should be able to take one segment, do the arithmetic by hand, and get the same answer.

A model that passes that test can be argued with, which is the point. A model that cannot is asking to be trusted, and trust is not a budget argument.

Testing the rates you are already using

There is a check available to any council with more than one year of roughness data, and it does not need a model at all.

Austroads suggests using a roughness time series to assess whether the intervention levels currently in force are working — and gives an example of a network where "the currently applied intervention levels do not appear to be holding the network in a stable condition", concluding that a change is probably required (Guide to Asset Management Part 5B, 2007).

That is a question your own data can answer about your own network, and it outranks any projection. If the network is drifting down under the rules you have, the rules are the finding.

Victoria's auditor-general reached the same conclusion about the state network from the other direction: around 48% of state roads were in poor or very poor condition in 2024 against around 39% in 2021, with the department advising government that funding was insufficient to maintain or improve condition (Maintaining State Roads, VAGO, September 2026). Those are state roads, around 15% of Victoria's network — but the method of noticing is the same one.

On machine learning

It works, and it is mostly not for councils yet.

The strongest published results come from state and national road authorities with deflection testing, mix records and construction histories spanning decades — data a small council does not have and will not have soon. More importantly, a model whose reasoning cannot be stated in a sentence breaks the one property a council most needs: the ability to explain a number to somebody who did not want to hear it.

There is a narrower use that is clearly worthwhile: statistics to flag an implausible survey reading for a person to check. That is a different job from producing the projection.

What to ask a vendor

  1. What rate are you applying, and where can I read it?
  2. Which inputs do you use, and what do you do when one is missing?
  3. Is traffic loading in the model, and is it axle-based or count-based?
  4. Which numbers in the output are measured and which are modelled — and does the output say which, where a councillor reads it?
  5. What happens to a road that has never been surveyed?

The fifth is the one that separates systems quickly.

How PaveKeep handles this

PaveKeep projects condition forward from the council's own data and a published method — the rates, thresholds and rules are on a page in the product, read live from the council's own configuration, including the ones it does not use. Heavy-vehicle loading changes the expected rate and the change is visible rather than buried. Where a number is modelled rather than measured, it is labelled as modelled wherever it is shown, and the two are never blended into one figure.

Climate is not an input. The structure for it exists and nothing consumes it, and the method page says so in those words.

PaveKeep is built for small and regional councils and is not yet in use at any council.

Sources