There’s a lot of excitement right now about what artificial intelligence can do with point clouds. Some of it is justified. Some of it is marketing noise.
Over the years I’ve worked with point clouds on industrial projects – refineries, chemical plants, combined heat and power stations – and helped roll out a platform for managing spatial data. That has given me a view of AI from both sides: from the data, and from the technology meant to process it. Here’s what I see.
What AI can do today: point cloud classification and segmentation
Let’s start with what AI genuinely does well.
Point cloud classification and segmentation is where machine learning has made the biggest progress in recent years. So what does that mean in practice?
A raw point cloud is a set of millions of points in space. Each one has XYZ coordinates and an intensity value – plus colour, if the scanner also took photographs. What it doesn’t have is a label. The scanner has no idea whether it has captured a pipe, a wall, a floor or a steel truss. That has to be worked out by a person – or by an algorithm.
Classification means assigning every point to a category: “this is pipework”, “this is a support structure”, “this is floor”. Segmentation goes a step further and groups points into individual objects: “this set of points belongs to one object”.
AI algorithms – especially deep learning networks trained on large industrial datasets – can do this far faster than a person. Work that would take days by hand can be done in hours. That’s a real time saving and a real gain in productivity.
At 3Deling we use these tools on our projects. As a result, classification and initial segmentation become the starting point for an engineer’s work, rather than a laborious stage they have to work through from scratch.
The key shift: a classified, tagged point cloud linked to documentation
I want to pause on one observation, because in my experience it changes how people think about the end product of digitalisation.
For years, the end product of a scan was understood to be a CAD or BIM model. The point cloud was raw material – something you “turned into” a model.
Today it’s increasingly clear that a classified, segmented and tagged point cloud, linked to technical documentation and 360° panoramas, is a valuable product in its own right. Not as a substitute for a CAD model, but as a data layer that opens up entirely new uses: searching for components by their attributes, integration with CMMS systems, automatically generated inspection reports, and tracking changes between scanning campaigns.
And this is where AI offers the most today – not in building CAD models, but in enriching spatial data and connecting it to documentation.
What AI can’t do yet: a CAD model is still an engineer’s job
I want to be precise here, because I hear far too many promises the market simply can’t deliver.
AI cannot, today, build an as-built CAD model in AVEVA E3D or an intelligent BIM model from the point cloud of an industrial plant on its own. Not because the algorithms are poor, but because the task is so complex, and the information so scattered, that no existing AI system can handle it end to end.
Modelling an industrial installation isn’t just about recognising geometry. It’s about interpretation. Is this pipe part of process line X or line Y? What’s its material specification? Which system does this valve belong to?
Then there’s a problem anyone who has worked on an existing plant will recognise: the 2D documentation often doesn’t match what’s actually on site. Working out where a discrepancy came from, and how to correct it, is something AI won’t be able to do on its own for a long time yet.
These questions call for engineering knowledge, an understanding of the technical context and – very often – cross-checking against P&IDs, material specifications and the history of modifications. No algorithm today has access to that knowledge in a way that could replace an experienced modelling engineer.
Anyone who tells you otherwise isn’t being straight with you.
That may change one day. But not soon. And not without a foundation of well-structured, classified data – which AI has to “learn” plant by plant.
What this means for your project
Getting to a fully working, fully deployed digital twin of a plant is a journey – usually a long and costly one. It takes effort not just from the technology provider, but above all from the organisation itself: changing processes, training people, and deciding what to digitalise and why.
But the journey can be broken into stages, and the benefits start early – well before a full intelligent model exists. A point cloud you can access in a web browser, linked to your P&IDs, is already a valuable operational tool. You don’t need to wait for a “complete digital twin” to start saving on site visits or responding faster when something breaks down.
If you’re wondering how AI fits into your digitalisation plans, try changing the question. Instead of asking “can AI do this for us?”, ask “how should we specify the scan deliverables so that AI can make the most of them?” The answer leads to decisions worth making at the very start of a project: what level of point cloud classification you need, which attributes should be tagged, and how the point cloud should link to your existing documentation.
The road to a digital twin starts with solid data. AI can shorten that road – but it can’t replace it.
Planning to digitalise your plant and want to know how to scope the scanning so you can get the most out of AI?
Get in touch – we’ll help you scope the project, from the control network to a classified point cloud that’s ready for the next stage.
