At a sawmill, hundreds of metres of wood pass through industrial laser scanners every minute, detecting surface irregularities with millimetre precision. Yet the question of the wood’s internal structure, which is crucial for a board’s strength, remains difficult to determine with certainty. Which boards are suitable for load-bearing beams, and which are better suited to panelling?
Anders Olsson, professor of building technology, holds up a board of spruce and points out visual clues to the quality of the timber. Each annual ring, knot and fibre affects the strength.
“At present, the only way to obtain a complete three-dimensional image of a board is through CT scanning. But that’s both slow and expensive, and not ideal for industrial sorting.”
Increase the value of wood
Anders has been interested in carpentry and buildings for as long as he can remember.
– Building technology is hands-on, and I enjoy sitting down to calculate and think. My research has always been driven by curiosity, but not only that – there has always been a planned and possible application in the background.
For more than 15 years, his focus has been on strength grading and increasing the value of wood material. His research group has long explored how to gain as much knowledge as possible about the internal properties of wood based only on what is visible on the six faces of a board. A good method must be both accurate and fast enough to cope with industrial reality, without damaging the material.
Developing a new AI method
It’s a challenge the researchers are well on their way to solving, by developing an entirely new method for strength-grading timber using artificial intelligence. The work gained momentum in 2019 when Osama Abdeljaber joined the group.
Timber was a completely new field for him, but he had developed software to detect damage in steel structures by analysing sound and vibration in the material. He saw an exciting challenge: just as with steel, the task was to understand the interior of a material without opening it up. AI could work here too.
“Wood is a natural material, and every board is different. That makes the challenge very complex, but also great fun. I saw huge potential in machine learning and was motivated by the chance to create valuable solutions”, says Osama, senior lecturer in building technology.
“We learn a lot from one another. I contribute my knowledge of wood, Osama his of machine learning, but we share a common foundation in mechanics. This collaboration has been a major step forward”, says Anders.
Generates virtual boards
The research is based on using the sensors already found in today’s sawmill sorting lines – cameras and lasers that scan all sides of each board on a conveyor – to calculate, with the help of AI, what the inside of the board looks like.
Lacking sufficient real-world data, the researchers developed a simulation programme that generates thousands of photorealistic virtual boards with precisely defined properties such as pith, annual rings and knots. These are used to train an AI model to understand the relationship between exterior and interior.
Teaching the AI new wood species works a bit like when humans learn a new language.
They are well on their way to success. Collaboration between the research group and the company Microtec, which develops scanners for timber sorting, has so far resulted in two patents.
“We’ve already succeeded in predicting the position of the pith and identifying annual rings in spruce boards. We’re now involving even more parameters, which produces more realistic boards and, therefore, more useful models for the industry”, says Anders.
A complete picture of the board’s interior
Rising timber prices and increasing demand highlight the need to use the best wood where it is most needed. This is why the project Detection and modelling of knots and fibre orientation in sawn timber – funded by the Knowledge Foundation and carried out in collaboration with Microtec, Ikea and Södra – plays such an important role.
“Companies such as Södra produce millions of cubic metres of sawn timber every year. With more precise and data-driven methods, we can make better use of the material, save money, and at the same time contribute to climate goals”, says Anders.
The long-term goal is to provide a complete picture of the interior of a board – including knots and fibre direction – using only optical surface scanning. This would make it possible to predict the strength of each board more accurately than is possible today and maximise value when thicker boards are sawn into thinner ones.
“With better knowledge, more boards can be used in the best possible way, allowing us to save timber material. We can also load wooden structures more efficiently because we know what each board can handle.”
Anders Olsson
Professor of building technology
Anders Olsson is a civil engineer specialising in civil and environmental engineering, with a doctorate in structural mechanics. His main interests are computational models for analysing wood materials and using such models to predict the properties of timber and other wood-based components.
GRANTS AWARDED SINCE 2010
SEK 25.7 million
Funders: The Knowledge Foundation, Vinnova, Formas, The Södra Foundation for Research, Development and Education, Centrum för byggande och boende med trä, The Nils and Dorthi Troëdsson Research Foundation
Osama Abdeljaber
Senior lecturer in building technology
Osama Abdeljaber is a civil engineer with a doctorate in building technology. His research focuses on AI and its applications in building technology, such as identifying properties and defects in wood, monitoring building health, and detecting structural damage.
Like learning a new language
Spruce is the most common wood species in the Swedish construction industry and a natural starting point for the project. But ambitions extend further. The next step is to adapt the model for pine and eventually for deciduous species such as birch, beech and oak, which differ more in structure.
“Teaching the AI new wood species works a bit like when humans learn a new language. For each language you already know, it becomes easier to learn the next. In AI, this is called transfer learning. Each time you adapt a model to a new domain – such as a new wood species – you generally need less training data. If we can solve this challenge with spruce, the development prospects look very promising”, says Osama.