AI cuts fluid simulation time fifteen times

Researchers in Osaka have developed an AI model that performs complex fluid simulations in minutes rather than hours, potentially changing offshore engineering while maintaining high accuracy. This advancement could accelerate the development cycle of offshore technology and implement real-time monitoring systems that were previously considered computationally impossible.

Traditional particle-based fluid simulations are critical to predicting wave behavior in marine applications and often require extensive computing resources. The new Graphic Neural Network (GNN) method developed by the Metropolitan University of Osaka reduces the computation time from about 45 minutes to just 3 minutes while retaining the simulation quality.

“AI can provide excellent results for a specific problem, but it often struggles with different conditions,” said Takefumi Higaki, Graduate School of Engineering and principal researcher at Metropolitan University, Osaka.

Breakthrough solves the notorious challenge of AI in generalization. Previous fluid dynamics machine learning models have been struggling when faced with different situations than training data. The Osaka team systematically evaluated a variety of methods to create a model that can handle various fluid phenomena with consistent accuracy.

For investors in the monitoring computing technology market, advances in computing fluid dynamics (CFD) are being integrated with multiple high-growth areas. According to Allied Market Research, the global CFD market is valued at around 2023 and is expected to reach USD 5.3 billion by 2033, with an annual growth rate (CAGR) of 7.2%. As the Globenewswire report highlights, other major forecasts will bring the 2030 markets close to $4.2 billion, with annual growth estimated at ranges from 6.8% to 9.5%, depending on sources such as Technavio, Digital Engineering 24/7 and Datahorizzon Research.

Marine engineering is a particularly important part of this ever-expanding market. CFD changes the design and optimization of ships, offshore structures and propulsion systems by providing accurate simulations of fluid dynamic performance, fuel efficiency and operating stability. Applications in this field are demonstrated by MR-CFD, Regalia Marine and Flow-3D. Adoption is accelerating as the industry seeks to address challenges such as fuel efficiency, emission standards and operational safety.

Meanwhile, AI-accelerated hardware is rapidly evolving to support physical simulation. As stated in this Nafems article, technologies such as AI acceleration solvers and specialist processors (especially NVIDIA GPUs) are promoting significant improvements in simulation speed and efficiency, where analysis time is sometimes reduced from hours to seconds. NVIDIA announced the open source physics engine for robotic simulation Newton announced how these advances can open new opportunities for professional hardware in engineering and scientific applications, further improving the performance and scalability of CFD and related technologies.

The study also had an impact on renewable energy development. Offshore wind turbine design and tidal energy systems require precise fluid modeling to optimize placement and energy capture. Faster simulation capabilities can compress the development timeline of these technologies while improving their economic viability with more accurate performance predictions.

What distinguishes this study from previous efforts is its methodical approach to improving generalization. Rather than focusing on speed, the team systematically analyzes which features are crucial to the accuracy of different simulation conditions. Their graphical neural network maintains high accuracy even when applied to fluid scenes not included in their training data.

The ability of this model to handle larger time steps than conventional methods is particularly noteworthy. Although traditional computational fluid dynamics (CFD) face stability problems when using larger time increments, AI models remain stable and accurate, with the time steps 10 times larger than their training data.

“Faster and more accurate fluid simulations may mean significant acceleration in the design process of marine and offshore energy systems,” Higaki notes. “They also enable real-time fluid behavior analysis, which can maximize the efficiency of marine energy systems.”

For offshore technology developers, the implications are huge. Hull designs are often extensive fluid dynamics testing before building physical prototypes. Speeding up this process from hours to minutes can translate into weeks or months saved during design iterations. Likewise, offshore platform designers can evaluate more design changes in less time, which can lead to a more optimized structure.

Although the current research focuses on two-dimensional fluid simulations, the team plans to extend its approach to more complex three-dimensional scenarios. They also explored the potential of directly outputting training models from experimental data rather than traditional simulations, which could further augment the applicability of the real world.

As computing resources are increasingly restricted relative to simulation requirements, methods that maintain accuracy while greatly reducing processing requirements may find rapid adoption among industries. This study represents an important step in enabling a wider range of applications and users to access complex fluid dynamics.

The study was published in the January 2025 Applied Ocean Research.

Related