Working with the Tor Vergata University Rome on Machine Learning to Improve the Efficiency of Fluid Dynamics
Study validates the effectiveness of Cogisen’s machine learning model to measure fluid dynamics
Machine Learning, Cogisen
Fluid Dynamics is around us every day. Any flow moving fast enough or with a small kinematic viscosity generates turbulence. Turbulent flows, like atmospheric and oceanic circulations or flows around vehicles, develop extremely complex dynamics coupling structures over a large range of scales, and with chaotic behavior, which makes them unpredictable. For engineers this creates significant design challenges as they have to find aerodynamic data that best replicates the conditions in the outside world. Today, scientists rely on experiments such as wind tunnels or numerical simulations performed on the largest supercomputers in the world, an approach known as computational fluid dynamics (CFD). Both approaches offer invaluable information, but also have limitations. For instance, when turbulence increases, simulations must find a way to reflect these ever-growing scales. They struggle, though, to reproduce all flow conditions as observed in nature and physics experiments, as they are limited by the computational power of the latest hardware. A more practical approach is the use of wind tunnels, but even though they can reflect more variants, they cannot perfectly replicate the outside world.
It is no surprise, then, that market research suggests there is going to be significant investment in these tools in the coming years, as engineers and academics seek to improve their effectiveness. The CFD market is growing significantly with IMARC Group expecting it to reach a value of US $3.1 billion by 2024. There is demand from diverse industries including aerospace, defence and automotive for more efficient fluid dynamics analysis, as the advent of unmanned aerial vehicles (UAVs) and autonomous and electric transport create new design requirements. Similarly, the market for wind tunnels is going to reach $2.88 billion by 2023 according to Markets and Markets.
The Cogisen Cognitive Laboratory understands the size of the challenge for engineers and has undertaken research with the Tor Vergata University in Rome to create a more efficient model for analysing fluid dynamics data, using machine learning algorithms. The goal is to create an approach, which can be applied in academic studies and is relevant to commercial requirements to quickly deliver accurate results. In sectors such as Formula One, this will be invaluable for designers, who are constantly using fluid dynamics to adjust and change the shape of their vehicles to secure any possible competitive advantage.
In the last few years, the machine learning focus in the technology industry has been on Convolutional Neural Networks (CNN), which have been proven to boost classification performance in several domains. Mostly, CNN applications are restricted to expressing input data in two-dimensional form and lack a scalable way to take advantage of the richer representation offered by 3D or 4D formats. Some efforts have been made to develop 3D or 4D CNNs architectures, but usually these configurations require a very high computational cost in terms of dedicated hardware and training time. Applying CNNs to turbulent flow classification is similar to any other image classification problem, but finding the best strategy to classify the input data changes, because of the complexity of the input data. Given the representation limitation of CNNs this is challenging when analysing turbulent flows, because it suggests they cannot take advantage of the underlying physics identified by richer representations.
In our work we want to use machine learning to infer the parameters of turbulent flows moving within a rotating system, using their velocity fields as input data generated by numerical simulations. The system external rotation has non-trivial effects on flow dynamics and it is a very common condition in nature, so it is important to interpret this data to find ways to improve engineering design. Indeed, when the angular velocity (Ω) is large enough, the flows develop large scale intense vortical regions, which are completely absent when the system is still or slowly rotating. The goal of our collaboration is to train a model able to infer different values of Ω by recognising key different spatio-temporal features in the input data.
(Volumetric representation of a system moving at different angular speeds. After a critical velocity there is the formation of vortices in the fluids. Colors represent the energy magnitude.)
The University team provided Cogisen with simulations to test against its artificial intelligence (AI) algorithms; in parallel, the team led by Dr. Buzzicotti was responsible for using the same simulations to feed CNNs. Preliminary results show CNNs are able to discriminate systems rotational speed among three general conditions: below the critical speed, around the critical value and above. However, while the number of speeds CNNs are able to detect will certainly increase, we were able to prove that CNNs were computationally demanding and it was difficult to interpret the results. This is because CNN models require huge amounts of data to run analysis and the results emerge from a “black box,” making it nearly impossible for researchers to understand why – and how – the neural networks have reached their conclusions. As a result, it is hard to foresee a short term use case of CNNs to help engineers examine the impact of turbulent flows on the performance of their designs.
We are delighted to say that our AI technology has already been able to identify four different speeds (one below, one around the critical speed and two above), profiting from the easy blend of time and volumetric information that is more complicated with CNNs. The mean accuracy reached in the classification was 99% and perhaps more interestingly than the performance, the Cogisen AI algorithms proved they could learn to develop a model from the data they were studying, ordering the fluid dynamic data proportionally to the speed and the energy of the system. This outcome will enable engineers to design with greater accuracy and efficiency, because it gives them a much more 3-dimensional view of the speed of vortices and turbulence, and their evolution in time.
(Output of Cogisen model. The results on the right, under TEST, are obtained on unseen data.)
We can deliver this outcome, because we are building machine learning models in the frequency domain using an approach to machine learning called cognitive modelling. It is inspired by the way the human cognitive mind identifies patterns to associate with objects and recognise them. The cognitive mind recognises objects with much greater efficiency than machines, so by taking this inspiration we are able to significantly reduce the amount of data required to train the application. For this study our algorithms were trained on a laptop CPU in less than an hour, which is a huge contrast to the vast amounts of data and time that CNNs require.
This approach is also completely transparent for those wishing to understand why and how our algorithms delivered these results. What makes our approach even more effective is that we are able to use a single model, which adapts to all four velocities. This ensures the output data matches the rotational speed of the input data and the energy organization within it. As such our algorithm can identify a clear law between the input and output data, simplifying the training process.
All in all, the Tor Vergata team has been impressed with the outcomes and we are now moving to develop a peer-review paper that we will publish later this year. We are also planning to expand this model to include entropy. Once we are able to integrate entropy with analysis of energy we will engage commercially with automotive and aerospace engineers to evaluate how our technology can deliver more effective analysis of fluid dynamics.
Silvia Colabrese is working on machine learning based data modelling at Cogisen and holds a PhD in Computational Vision, Automatic Recognition and Learning from the Italian Institute of Technology (IIT) and the University of Genoa.
Modelling how we perceive video