Safer Bridges with Machine Learning

The challenge

A bridge collapse reverberates around the world. The random but catastrophic nature of such an event looms large in the public consciousness, and the flow-on effects for an affected region’s social, emotional and economic health can be massive and devastating. How can we use the latest engineering and computational techniques to design and maintain safer, better bridges?

In this article

  • How scour causes bridge failure
  • How machine learning can be used to monitor scour to provide early warning systems for bridge damage and failure
  • What machine learning means for the future of bridge design and monitoring
  • How we define the difference between machine learning and artificial intelligence

This project was funded as part of our 2018 Global Research Challenge, an annual invitation for our people, clients and collaborators to present and develop an idea they think will help tackle one of the world's most intricate problems. This project was proposed under the theme, 'Artificial Intelligence: Designing and Engineering in the Built Environment.'

Twice a year, the Swansea Bridge—which connects New South Wales’ Central Coast to Newcastle—must be partially closed for maintenance. For three days every May and October, one lane in each direction is shut and the 27,000-vehicle daily capacity is halved. Commuters are asked to find alternative routes. Vacationers prepare for delays. Boats are unable to cross through the bridge’s navigation channel, closing the only gateway available for large boats and yachts between Lake Macquarie and the Pacific Ocean. While there are many reasons to close a bridge for maintenance, one is tidal scour. Fast-moving water erodes sand or gravel from bridge abutments or piers, making them unstable.

In 2012, the US government spent $17.5 billion on bridge construction. From that, a staggering $16.4 billion went to rehabilitate or replace existing bridges. Again, scour at the piers of bridges passing over waterways was shown to be the main culprit of bridge collapse. The personal and financial costs of this natural phenomenon are substantial. If we ever hope to design for bridge scour, we need to get better at identifying and predicting how and where it happens. 

We can get some of this data by monitoring bridges over time, which can help engineers design more reliable bridge structures, and help the infrastructure owners improve their asset management plans. Unfortunately collecting and processing all this data is a slow and arduous task. For the past two decades, enormous research effort has into monitoring bridges. In Australia, the Australian Network of Structural Health Monitoring helps coordinate and integrate efforts for better development and application of monitoring techniques in Australia. Several university and governmental agencies are part of this network, including the Queensland Department of Transport and Main Roads, New South Wales Roads and Maritime Services, Queensland University of Technology, the University of Adelaide, the University of Melbourne and University of Technology Sydney.

“Everyone knows scour is a problem,” says Negin Yousefpour, an engineer from the US who recently relocated to Melbourne to work on the Metro tunnel project. Negin believes we can use modern computing techniques like machine learning to comb through all the data we’ve been collecting and derive insights about how to detect scour and predict failure before it occurs.

“There is bridge monitoring data out there to help build a picture of various issues over time, but very few studies try and use that data to find solutions using the emerging computational power we have access to nowadays. This data can help us predict bridge damage, improve public safety and limit economic loss. Our study brings together some of this available data and process it using machine learning techniques, with a goal to develop a robust and cost-effective early warning tool for bridges.”

Bridge scour occurs when fast-moving water erodes sand or gravel from bridge abutments or piers, making them unstable. Tidal forces are often a main contributor to bridge scour.

Now is probably a good time to explain what Artificial Intelligence (AI) and Machine Learning actually are. While their exact definitions depend on who you ask, we think of Artificial Intelligence as a broad concept that describes machines or robots doing tasks or making decisions that appear ‘intelligent’—operating an autonomous car, for example. Some purists claim that Artificial Intelligence will only be obtained once robots become sentient, but that’s still some time away we hope! Machine learning is less glamorous but arguably more applicable in our engineering design world. We define it as giving a computer access to data and letting it learn for itself. The more data given to a machine, the quicker and more powerful its ability to derive insights becomes. Think of Facebook’s facial recognition algorithms. Remember years ago, when it still had problems telling you and your sister apart, or differentiating your two blond-haired friends? Billions of images later, it can now pick out a blurred face in the back of a crowd. 

We’d like to train our algorithms to perform the somewhat similar task of looking for failures, flaws, or weak points in our infrastructure. Unfortunately, we don’t have quite the same organised database as Facebook. But we do have millions of concept, scheme and detailed designs, models, performance and budget data from the projects we've run around the world for the last sixty years. We’re working on it.

For the past few months, Negin has been working with Texas A&M University, various US Department of Transportation bridge engineers, and the Transportation Research Board Committees to see how algorithms can be used to identify bridge scour. They also see AI and machine learning as having huge potential benefit when it comes to making forecasts on bridge performance trends, because it will greatly reduce the cost of monitoring, repair and retrofitting measures.

“Our whole industry is excited about the potential of AI and machine learning,” says Alex Sinickas, who leads our Research Program. “But the real trick is to match its powerful methods, which require massive datasets, with a problem that we really need to solve. Usually our machine learning projects run into issues where we have a problem but insufficient or low-quality data, or vice versa. In Negin’s case, we think we might have found a good match.” 

Over the last few months, Negin has been scouting for extra monitoring data from Australia, Asia and the UK to add to the American dataset. The data usually come from scour instrumentation as well structure health instrumentation, such as strain gauges and accelerometers. In many countries, it’s collected by government departments responsible for Transportation. For example, in the US, the Department of Transportation and the Federal Highways Administration are leading the collection of instrumentation data from bridges in various states.

Machine learning can be applied to all sorts of engineering problems of course. One of our first forays into using it, not just for monitoring, but for prediction, was for the Crossrail project - digging 42 kilometres of bored rail tunnels beneath Central London. 

Using the current dataset, Negin’s team has started to develop and train Artificial Neural Networks to identify the patterns indicating bridge malfunction/failure due to scour at bridge piers. Also known as neural nets, these are computing systems inspired by the circuits of neurons and synapses within our brains. These digital ‘brains’ learn to perform tasks by processing large volumes of information and teasing out trends. Think of a baby learning her first words. How quickly your neural net can learn comes down to the volume and the quality of the data you feed it. The more reliable the information that goes in, the more accurate its findings become.

Once Negin and her team feel their algorithms have been trained and properly tested to recognise bridge scour based on real-time data, which means the algorithms can successfully identify where a “known” failure has occurred, they can then be set free to monitor existing bridges to identify scour in real time. Easier said than done, of course, Negin expects the development process to be complete in the next two years.

The output, they are hoping, will be well worth that wait. Learning from past data, Negin’s algorithms will be useful for predict future damage. Her team's model may also reduce the time and cost associated with bridge monitoring and repairs and provide a greater understanding of bridge malfunction patterns.

“What we are doing is unique” Negin says. “No-one else has brought real time data together in this way and applied ML solutions to predict bridge malfunction.”

Nathan Perkins, a programmer in our Melbourne office is involved in several machine learning projects, as well as research into robotics in construction with local contractors.

Negin’s project is one of many where we have applied machine learning to performance data. You can read about how our facades team is combining algorithms with drone imagery to speed up and de-risk the inspections or facades in Issue 1. While each of our machine learning projects focuses on a specific problem, when they are aggregated, we see a fundamental shift in our ability to improve the safety, design and performance of infrastructure.

In the meantime, the algorithms Negin and her team have developed are hungry for data. They’re always looking for more. If you or your organisation has access to bridge monitoring data and you are interested in contributing and learning from this research project, please contact her via the link below.


  • A plethora of bridge instrumentation data exists. We can use machine learning algorithms to sift through this data to better understand how and where scour occurs.
  • Once we understand bridge scour a bit better, we can use these algorithms to monitor bridges against scour in real time.

This story was written by Liz Uhlmann, as part of the Research Review series. The series is produced by the Arup Australasia Research team; Alex Sinickas, Bree Trevena and Jeff McAllister with contributions from Sheda and Noel Smyth.

Lead Arup Researcher

Dr Negin Yousefpour
Negin is a geotechnical engineer in our Melbourne office.

Ask Negin about:

  • Current bridge scour monitoring techniques and instrumentation.
  • Training artificial neural nets to recognise scour early on.
  • How your data can contribute to building better, safer infrastructure.


Professor Jean-Louis Briaud
Dr Briaud is a professor of Civil Engineering at Texas A&M University and is our main academic collaborator.

Research TEAM

Steve works with our Advanced Technology and Research Group, and is the Director for this research project.
Ben is an engineer with the Advanced Technology and Research Group in London.
Graduate analyst, Melbourne office
Dr Steve
Steve is from our Advanced Technology and Research team in London.
Hristo is a bridge scour expert in our Advanced Technology and Research team in London.
Dr Pablo
Pablo works with our Advanced Technology and Research team in New York.
Dr Matthew
Assistant Professor Yarnold is an expert in bridge engineering at Texas A&M University and one of our academic collaborators.
Dr Robin
Dr BeeBee is a hydrologist with the Alaska Science Center and one of our industry collaborators.
Dr Ken
Dr Fishman is a geotechnical engineer, committee member for the Transportation Research Board, Chair of the TRB Bridge Foundation Committee and one of our industry collaborators.
Dr Sharid
Dr Amiri is a Geotechnical Engineering Specialist with the California Department of Transportation, Chair of the TRB Geo-seismic Committee and one of our industry collaborators.

Have a problem or a project?

We work with industry partners, governments, universities, startups and community organisations. We do this through research partnerships, and as consultants and facilitators for foresight, research, storytelling and technical writing workshops.

research with us