BIM clash detection: triaging real clashes from tolerance noise

· Postlight Research

Run a hard clash test across the federated model of a mid-size building and the first report rarely comes back as a short list. It comes back with thousands of hits, and most of them are not decisions anyone will ever make. This note explains where that noise comes from, how experienced MEP and structural coordinators decide which element moves, and what an automated triage system has to learn before its output is worth reading.

What a clash test actually computes

Navisworks Clash Detective, and the IFC coordination tools that work the same way, test geometry, not intent. A hard test reports any pair of elements whose triangulated meshes intersect, filtered by a penetration tolerance: set it to zero and every overlap counts, set it to 5 mm and shallower overlaps drop out. A hard (conservative) test also reports faces that only touch. A clearance test reports pairs that come closer than a set distance, which is how insulation, access space and hanger zones are checked. A duplicates test finds the same element modelled twice in the same place.

Each test is exact about geometry and silent about meaning. The usual sources of hits that nobody acts on:

The clash count therefore measures model density as much as coordination quality. Grouping, rule sets and selection filters reduce the pile, but someone still has to read every group and decide.

How coordinators actually decide

Experienced coordinators do not resolve clashes one hit at a time. They apply a priority order, often called a yield hierarchy, which says which system moves when two systems want the same space. The order varies by project and contract, but the logic behind it is consistent: the element that is hardest, most expensive or least able to move keeps its position.

Illustrative: a common yield hierarchy, not a rule for every project.

Priority System Why it usually holds its position
1 Primary structure Columns, beams and slabs are sized and placed by structural analysis; moving one reopens the design.
2 Gravity drainage Soil and waste pipes run at a fixed fall, so a change at one point moves every point downstream.
3 Large ductwork Big rectangular ducts have little room to offset, and every extra fitting adds pressure loss.
4 Pressurised pipework Water lines under pump pressure can rise and drop with fittings at modest cost.
5 Cable tray and conduit Electrical containment bends easily and is usually routed last.

Inside that order, coordinators weigh facts the clash report does not contain: how much fall a drainage run has left before its outfall, whether a beam can take a web penetration and where along its span, whether a duct can be flattened to an equal-area section to pass under a beam, and what a move does to the ceiling height. A clash between a sprinkler main and a cable tray is settled in the meeting. A clash between a drainage stack and a transfer beam can take a structural engineer, a revised penetration detail and a week.

That judgement is the expensive part. The geometry test runs in minutes; the ruling does not.

What an AI triage system has to read

A system that reads only the clash report sees what the coordinator sees in the first minute: two element identifiers, a point and a penetration depth. That is not enough to rule. To separate real clashes from tolerance noise, and to propose which element yields, a system needs inputs that describe the decision, not just the collision:

Learning from paired revisions is the central move. A coordination history is a record of many small engineering decisions with their outcomes attached. From that record a system can learn that a 3 mm overlap between a wall and a slab is never moved, that an insulated pipe touching its own hanger is never an issue, and that a drainage run crossing a beam ends in a penetration detail or an upstream reroute, almost never in the beam moving.

Evaluating automated clash triage on a real project

Clash triage accuracy is reported in many ways, and few of them survive a live project. A useful evaluation runs on the team's own model, with its own history, and measures the system against rulings the team's coordinators already made.

  1. Freeze a past coordination round. Take the federated model from before a round of meetings and the model from after, with the clash report and the issue log in between.
  2. Hide the outcome. Give the system only the before state and the clash report.
  3. Score triage apart from resolution. First: did it separate actionable clashes from noise as the coordinators did? Second: for the actionable ones, did it pick the same yielding element and propose a move of similar size and direction?
  4. Count the clashes it creates. Rerun the clash test on the proposed after state. A resolution that clears one hit and creates three is not a resolution.
  5. Read the disagreements by hand. Some will be errors. Some will be calls the coordinators made under schedule pressure that a second look would change.

Illustrative: a scoring sheet for one coordination round. Fill it with your own counts.

Measure What it tells you
Noise correctly dismissed How much reading time the system saves
Real clashes correctly kept Whether it is safe to trust the dismissals
Same yielding element as the coordinator Whether it understands the hierarchy
New clashes created by its proposals Whether its moves respect the rest of the model

The second line matters most. A triage system that dismisses one real clash in a plant room, a clash that is then built on site, costs more than all the reading time it saved. Weight the evaluation accordingly and look at recall on real clashes before anything else.

Where this leaves coordination teams

Clash detection tools are not the bottleneck; they find geometric conflicts quickly and reliably. The bottleneck is the ruling: deciding, for each real conflict, which element yields and how. That decision can be learned, because coordination teams have been recording it for years in revision pairs and issue threads without calling it training data.

Postlight is building models that learn that ruling from an engineering team's own project history. Read how we approach automated clash resolution for engineers, or see how Postlight keeps a model current during construction in BIM progress monitoring.