Data & analytics, Research & development, Technology & innovation

Pioneering machine learning for next-gen water modelling: fast physics-guided surrogates

PhD student Freja Petersen has joined DHI to explore how machine learning can speed up hydrodynamic modelling.

In a world where speed and accuracy are critical for addressing environmental challenges, PhD student Freja Petersen is pushing the boundaries of hydrodynamic modelling. She is doing her industrial thesis at DHI and Technical University of Denmark (DTU), and her research focuses on using machine learning to create fast, physics-guided surrogates of MIKE water models. We sat down with Freja to learn more about her project and its potential impact on the water sector and beyond.

Q: Can you tell us about the focus of your PhD?

The overarching topic of my PhD is ‘Fast physics-guided surrogates for flexible mesh water models’. My goal is to develop machine learning-based surrogate models to supplement DHI’s coastal ocean models, enabling faster simulations while keeping as high an accuracy as possible. While DHI’s current models are robust and physically grounded, they can take hours to run. I’m working to create models that deliver almost the same outputs up to 1,000 times faster. This opens doors to real-time forecasting, data assimilation, ensemble modelling for uncertainty analysis and risk-based estimates – things that are challenging with current computational constraints.

Q: Why is speed so critical in this context?

In scenarios like storms or heavy rainfall, decision-makers need rapid insights to respond effectively. Fast-running models could predict the impacts on local areas, aiding evacuation efforts and mitigating risks. Beyond that, faster models benefit DHI’s internal operations and clients by speeding up tasks like calibration and improving overall performance with real-time systems.

Q: How does machine learning come into play?    

Machine learning (ML) has enormous potential to complement traditional hydrodynamic models. By integrating data-driven methods with physics knowledge, we can overcome some of the limitations of purely data-driven approaches, like poor generalisation across domains. My focus is on scientific machine learning, a rapidly evolving field that bridges the gap between physics and ML. At this point, it is important to stress that these surrogate models won’t replace existing physics-based models. On the contrary, they rely on their data for training.

Q: What are the benefits of leveraging ML for hydrodynamic modelling?

ML can drastically reduce the time it takes to simulate complex systems. For instance, in ensemble modelling, where multiple simulations are run to account for uncertainties in input data, speed is crucial. Faster models enable more comprehensive risk assessments and better decision-making. Additionally, I hope my work will not only advance DHI’s offerings but also contribute to the scientific community’s understanding of ML applications in water environments.

Q: What about artificial intelligence? How does it relate to your work?

AI is a broad term encompassing everything from large language models (LLMs) to autonomous systems, whereas ML is a subset that also includes more traditional techniques like regression. My work falls under scientific machine learning, which I wouldn’t classify as AI since AI is often commonly associated with ChatGPT or LLMs in general. Rather, my project aims at constructing advanced regression methods that have a strong connection to physics and scientific computing, and which are less prone to the biases and low reliability we know from LLMs. That said, AI-driven advancements like foundation models – which can be fine-tuned for specific cases – might inform future iterations of this project.

Q: Why did you choose DHI for your PhD?

During my Master’s, I delved into the scientific machine learning field, and when it came time to look for opportunities, DHI stood out. Their focus on sustainability and climate challenges aligns with my ambition to make a meaningful impact. Plus, their expertise in hydrodynamic and environmental modelling makes DHI the ideal partner for this project. With funding split between DHI and Innovation Fund Denmark, I feel supported in bridging academia and industry.

Q: Where do you see the future of hydrodynamic modelling heading?

I believe ML will transform simulation practices while also driving demand for more simulation. As the field matures, tools like these surrogates could become valuable complements to existing legacy systems. I hope my work inspires the development of efficient, accessible models that don’t require supercomputers for training – making them practical for a wide range of applications.

Q: Any final thoughts?    

There’s been a lot of hype around data-driven methods, but also growing awareness of their limitations. By combining physics knowledge with ML, I hope to strike a balance between innovation and reliability, creating tools that are not only faster but also trustworthy. It’s exciting to be at the forefront of this intersection, contributing to both DHI and the broader scientific community.

Freja’s work exemplifies the potential of scientific machine learning in water modelling. Stay tuned for updates as her research progresses. Freja can be contacted at frtp@dhigroup.com.

Click here to learn more about DHI’s application of AI and ML: DHI technologies | MIKE water modelling | Digital solutions