Fluid Intelligence: AI Redefines How We Simulate the Sea

Artificial intelligence (AI) is rapidly transforming our world. By automating repetitive tasks, optimizing workflows, and improving decision-making, this technology promises substantial gains in efficiency and productivity across industries. 

With this boost in productivity and efficiency, AI is projected to contribute trillions to economic growth and development worldwide. 

AI has also showcased tremendous potential in tackling complex challenges such as disease and climate change and driving innovation in various sectors to enable the creation of new products, services, and business models. 

An interesting application of AI has also been seen in fluid analysis. In machines, the testing of fluids like lubricants, coolants, and fuels is done in order to identify issues that can indicate potential problems or failures.

This allows for timely maintenance and repairs, which can then help prevent costly breakdowns and downtime. It also minimizes the need for major repairs and replacements and ensures that the machinery is operating at its peak performance.

For many years, fluid analysis has been rather long and cumbersome. But the advent of AI has made the entire process more simplified, efficient, and precise. 

After all, AI and machine learning techniques use massive data sets, learn from them, and then make predictions. The technology can take an asset’s full dataset over its lifespan into account, use multiple signals at once, and learn to adapt through feedback. 

Fluid analysis, however, goes far beyond oil in machines. In coastal and ocean engineering, fluid behavior plays a critical role in designing maritime structures, modeling shoreline changes, and even harnessing wave and tidal energy. 

Advancing Coastal and Marine Intelligence with AI

In coastal engineering, AI has made many improvements by addressing problems like sediment transport, shoreline dynamics, design optimization, coastal monitoring, and climate resilience.

An example of this was seen late last year, when researchers from the City University of Hong Kong utilized machine learning to enhance the accuracy of modeling the boundary layer wind field of tropical cyclones.

“We human beings are living in this boundary layer, so understanding and accurately modeling it is essential for storm forecasting and hazard preparedness.”

– Author Qiusheng Li

Because the air in this layer interacts with land, the ocean, and everything else at the surface level, modeling has been rather challenging. Despite traditional approaches using tons of data and running large numerical simulations on supercomputers, they still often result in inaccurate or incomplete predictions.

The latest study used an advanced physics-informed ML framework that requires a small amount of real data to capture the complex behavior of tropical cyclones’ wind fields, which contain information about the storm’s structure, intensity, and potential impact. The author Feng Hu said:

“With more frequent and intense hurricanes due to climate change, our model could significantly improve the accuracy of wind field predictions. This advancement can help refine weather forecasts and risk assessments, providing timely warnings and enhancing the resilience of coastal communities and infrastructure.”

Around the same time, separate research introduced an AI surrogate¹ to simulate the propagation of coastal tidal waves in an estuary for hindcast and forecast purposes. This approach accelerates simulations and incorporates a physics-based constraint to detect and correct inaccurate results.

By reducing the time cost of 12-day forecasting of traditional ROMS simulations to just 22 seconds, the research contributes to oceanographic modeling by offering a fast, accurate, and physically consistent alternative to traditional simulation models, particularly for real-time forecasting in rapid disaster response.

Earlier last year, a team of researchers also worked on improving the neural modeling² of Lagrangian fluid dynamics.

Based on that, the researchers enhanced both training and rollout inference of GNN-based simulators with varying components from standard SPH solvers, including viscous, pressure, and external force components.  The neural SPH-enhanced simulators then achieved better performance than the baseline GNNs, which, it noted, allows for considerably longer rollouts and better physics modeling.

Evolution of ML-Based Surrogate Models in Fluid Simulation

When it comes to fluid simulation, a common approach taken is the particle technique, where particles simulate the behavior of fluid flow. Some widely used examples include smoothed particle hydrodynamics (SPH), moving particle semi-implicit (MPS), or incompressible SPH.

However, these techniques need extensive computational resources, including processing power, time, and cost. In recent years, the need to simulate real-world fluids in every phase of engineering, from design, manufacturing, and development to verification, operation, and visualization, has grown, so computation time also needs to be reduced. 

Over the last few years, several ML-based surrogate models have been introduced to estimate fluid dynamics with a smaller computational cost.

This includes using machine learning to substitute particle methods and fast computation of Lagrangian fluid simulations, which involves tracking individual fluid particles and focusing on their trajectories and properties.

While ML can speed up Lagrangian fluid simulations, previous studies haven’t been able to validate the generalization performance of such surrogate models across diverse fluid behaviors. 

Then there’s the fact that most of these models have been validated under the CFL condition, much like traditional Computational fluid dynamics (CFD) methods, which limits their ability to significantly reduce computation time.

Moreover, the focus of such initial studies wasn’t on accuracy but rather on replicating fluid-like behavior in simulation environments.

So, subsequent studies made progress on surrogate models for SPH, which saw a gradual improvement in accuracy. Researchers used different methods for this, such as using SPH results as training data to perform Lagrangian fluid analysis with deep neural networks (DNN) and introducing graph neural networks (GNN) to learn the motion of fluid particles from SPH data, among others.

However, these didn’t aim to estimate pressure, one of the essential factors in understanding fluid mechanics and its interaction with structures.

So, the progression from there led to the recent trend of focusing on incompressible fluids’ pressure. For this, scientists accelerated solving PPE in MPS by using DNN. They introduced FGN or fluid graph networks (which leverage GNNs to simulate fluid dynamics) by using MPS as training data. They, however, didn’t confirm whether the estimated pressure reproduced the actual phenomena.

Most of these studies haven’t clarified the effect of different feature settings on the results either. But because the results likely depend on the feature settings, it is important to reveal which feature is essential for surrogate models to reproduce fluids.

So, a new study, published in Applied Ocean Research³, has presented a particle-based surrogate model that can be applied to larger time step sizes and different fluid phenomena.

The study presented three improved graph network-based simulator (GNS) versions, which had GNN learn the motion of fluid particles from SPH data. This includes GNS with pressure estimation (GNS-P), GNS with wall boundary nodes (GNS-W), and GNS with a combination of both (GNS-WP). 

In their study, the researchers demonstrated that pressure estimation is important for accurately predicting fluids and verified that the wall boundary nodes are vital in managing the moving wall boundary conditions. They also showed that GNS-WP was able to replicate the sloshing pretty accurately even when the simulation speed (the time step size) was 10x bigger than that of the training data.

The proposed method (GNS-WP) trained in the sloshing scenario, as per the study, can be applied to three different problems — hydrostatic, dam break, and free oscillation tests.

Click here to learn how the Venus Flower Basket is reimagining aero and fluid dynamics.

Faster, Smarter, & Scalable GNN-Powered Surrogate Model

The new ML-based fluid simulation model that reduces the computation time significantly without affecting the accuracy has been created by researchers from the Osaka Metropolitan University. This swift and high-precision method has potential usage in real-time ocean monitoring, ship design, and offshore power generation.

The AI-powered models are gaining a lot of traction in the fluid dynamics space thanks to making fluid simulations simple and faster. However, this technology has its own issues. 

As the lead author, Takefumi Higaki, who’s an assistant professor at Osaka Metropolitan University’s Graduate School of Engineering, noted:

“AI can deliver exceptional results for specific problems but often struggles when applied to different conditions.”

So, the team built the new model using graph neural networks (GNNs), a deep learning technology, to provide a tool that is consistently fast and accurate.

GNNs are a type of neural network architecture that processes and learns from graph-structured data. Graphs are data structures consisting of nodes, which are entities like products and people, and edges, which represent relationships between them. GNNs analyze large and complex relationships within the graph.

This neural network is used in social network analysis to understand patterns, predict user preferences based on interactions, model and predict the properties of materials, and identify potential drug candidates and predict drug efficacy.

In the latest study, a node is a fluid particle, while the edge is the interaction between those particles. 

The research team first determined which factors are important for high-precision fluid calculations. They compared different training conditions and then assessed how effectively their model could adapt to different simulation speeds and varied fluid movements.

The team found the results to show strong generalization capabilities across different fluid behaviors, significantly increasing speed and decreasing the time required for processing. 

Their technique has been reported to achieve an accuracy of the same level or even better than MPS, with 10x the speed on CPU and more than 200x faster on GPU. The study also noted that, despite training only using a powerful sloshing flow, GNS-WP was able to successfully reproduce both calm and static flows with a different wall boundary.

“Our model maintains the same level of accuracy as traditional particle-based simulations, throughout various fluid scenarios, while reducing computation time from approximately 45 minutes to just three minutes.”

– Higaki

With this achievement, the research is offering a scalable and generalizable solution for high-performance fluid simulation that balances efficiency with accuracy. More importantly, these improvements aren’t limited to the lab.

Higaki said:

“Faster and more precise fluid simulations can mean a significant acceleration in the design process for ships and offshore energy systems. They also enable real-time fluid behavior analysis, which could maximize the efficiency of ocean energy systems.”

Unlike other studies, this one detailed a step-by-step improvement of the particle-based surrogate models, which helps with their further development. 

In future work, the researchers also plan to address the challenges of insufficient physical consistency, management of unknown pressure, and extending the model’s use to complex and 3D problems. 

Given the proposed method’s potential to learn from experimental data in the real world, the team further aims to use this study as a foundation to recreate complex fluid behaviors whose governing equations aren’t known, like multiphase flows with discontinuous materials.

Innovative Company

Huntington Ingalls Industries, Inc. (HII -0.89%)

A major shipbuilder for the U.S. Navy, Huntington Ingalls Industries is constantly exploring improvements in fluid dynamics simulations to streamline vessel design and performance testing.

The company is building the next generation of smart defense and intelligence systems and is leveraging artificial intelligence (AI) to do that. By combining the power of cloud computing and edge devices with tailored software, HII hopes to make seamless human-AI teams the standard in future operations.

At HII, leading-edge AI and machine learning algorithms are developed, tested, and integrated to optimize and accelerate mission-critical systems and platforms.

The company’s advanced ML applications support a wide range of defense needs, including radio-frequency spectrum, automated analysis of imagery, cyber data, acoustic environment, and natural language for intelligence production.

HII also utilizes its deep domain and data knowledge to build ML for mission resiliency, operational readiness, and fleet sustainment under contested logistics. As an extension of its simulator-based training, it is involved in the development of ML-based operational decision aids.

In addition, the autonomous maritime platform developer has deployed deep learning (DL) AI with transformer-based architectures for precision signal search in massive volumes of highly cluttered data.

HII exploits and fuses data across different modalities, including RF spectral signals, geospatial imagery, and natural language media to improve accuracy. It also employs advanced NLP technology based on DL and ML to help with information sorting and connecting for global missions.

HII’s AI-powered digital twins, meanwhile, enable testing, validating, and saving resources from shipbuilding to fleet sustainment. Its advanced autonomy suite, Odyssey, transforms any vehicle into an intelligent robotic platform that enables multi-vehicle collaborative autonomy, health monitoring, sensor fusion, and AI-enabled perception. 

So, the company makes extensive use of the latest technology to boost its productivity and optimize its processes. However, HII does acknowledge that the inherent uncertainties of AI could lead to operational inefficiencies and competitive harm, especially if their tools prove to be inadequate.

Financially speaking, the all-domain defense provider has a market cap of $7.25 billion with its shares, as of writing, trading at $184.95, down 2.13% so far this year. With that, its EPS (TTM) is 13.96, the P/E (TTM) ratio is 13.25, and the ROE (TTM) is 12.56%. As for dividend yield, it’s a nice 2.92%.

When it comes to company financials, for Q4 of 2024, the company reported a revenue of $3 billion compared to $3.2 billion from the same quarter in the previous year. This drop was due to lower volume in all segments. Non-GAAP operating income also declined from $330 million in 4Q23 to $103 million, which was driven by lower performance at Newport News Shipbuilding. 

Diluted earnings per share during the quarter were $3.15, while it had a backlog of $48.7 billion at the end of the year.

For the full year, HII’s revenue was $11.5 billion, a slight increase (less than 1%) from 2023 due to higher volumes at Ingalls Shipbuilding and Mission Technologies. Diluted earnings per share were $13.96. 

Net cash from operating activities was $393 million while free cash flow was $40 million, a big decline from $970 million and $692 million respectively in the previous year.

In 2024, the company also reported achieving critical shipbuilding milestones, which included delivery of amphibious transport dock Richard M. McCool Jr. (LPD 29) and Virginia-class submarine New Jersey (SSN 796). HII’s Mission Technologies segment, meanwhile, secured awards with a total contract value of over $12 billion.

“We continue to make progress on ships put under contract pre-COVID, and are working diligently with our customers to put over $50 billion of new work under contract. Mission Technologies continued its strong track record of top line growth and margin expansion. We enter 2025 focused on our mission to deliver the world’s most powerful ships and all-domain solutions in service of the nation.”

– CEO and president Chris Kastner

Conclusion

AI-driven fluid simulation models have been helping advance the field of fluid dynamics for some time now. However, the latest AI model, which uses graph neural networks, has demonstrated impressive results. It not only showcases a high potential to bridge the gap between accuracy and efficiency but also enables faster, real-time applications in maritime industries.

This technology can prove promising in accelerating the design and testing of ships and offshore infrastructure and optimizing renewable ocean energy systems. The AI-based fluid simulation model can contribute to cleaner and smarter ocean engineering, leading to a thriving future!