Realtime CFD: the advent of AI/ML and Operational CFD Digital Twins!

Very rarely in one’s career does something happen which is both a paradigm shift and a game changer at the same time, related to the industry that you’ve devoted most of your career to – in my case Computational Fluid Dynamics (CFD). But, recently, I have seen something I thought I would never see in my 30 plus years career - a viable approach to Realtime CFD (and indeed real-time CAE) that delivers repeatable engineering level fidelity suitable for CFD analysts.?

Those of you who attended the Computer-aided Engineering (CAE) industry’s biennial global gathering way back in 2011 - the NAFEMS World Congress in Boston - will remember that Dr. John Parry and I, both working for Mentor Graphics at the time, presented a thought provoking paper on the theme of Historical Trends within the CFD industry with ‘Realtime CFD’ being identified as the ultimate ‘Holy Grail of CFD’. We posited that with real-time, engineering-fidelity CFD, a true democratization of CFD usage should occur in every industry for every engineer and plant operator in the world that could in theory benefit from CFD engineering simulation. I would also argue that it is apparent to many companies today that real-time CFD or CAE will ultimately also allow CFD or CAE to ‘shift right’ and be deployable beyond the R&D department into the manufacturing and production departments of all manufacturers around the world. And, CFD could also be deployed even later into the process or product installation/launch phase and into the day-to-day operation of actual products and processes until their ultimate retirement, disposal or recyclability. This use of CFD all the way through the complete lifecycle of a product or process from concept to circularity could be termed the ultimate ‘Digital Twin’ and would include the key ‘Operational Digital Twin’ or ‘Lifetime Digital Twin’ phase where huge benefits could be obtained – see the diagram below for a gas turbine related schematic of what I mean.

John Parry and I examined within our NAFEMS 2011 paper the historical trends visible in the CFD industry at the time and the apparent ‘game changers’ then visible on the horizon such as HPC, Cloud, CAD-embedded CFD, appification of CFD, 1d-3d CFD, and Lattice-Boltzmann CFD to name but a few. Since then, CFD has still not really been democratized very much in my opinion – we have seen attempts at “discovery” level CFD tools that have tended to produce results that were way off acceptable engineering accuracy levels, and we have seen all sorts of new SaaS CFD Apps appearing. Well, one technology that was off our radar screens ten years ago was Artificial Intelligence (AI) and Machine Learning (ML) although another Mentor Graphics colleague, Dr. Ivo Weinhold, and I in a separate 2017 White Paper on the topic of the Democratization of CFD touched on this issue in the form of talking about “intelligence” in CFD simulation with respect to greater CFD usage...

...all seemed to trundle on in the CFD world pretty much as it was I would say over the last 10 years until Hexagon acquired CADLM in April 2021, a fast growing Paris-based AI (Artificial Intelligence) and ML (Machine Learning) for CAE simulation company run by Professor Kambiz Kayvantash, a long-time veteran of CAE simulation, who has been working in the AI/ML field for the last 20 years. Hexagon has subsequently taken CADLM’s tried-and-proven general-purpose AI/ML algorithms for CAE embedded within a product suite called ODYSSEE CAE and recently released another product called ODYSSEE A-EYE to extend further into AI applications. These products really both include and extend beyond trusted design space techniques like Design of Experiment (DoE), Process Integration and Design Optimization (PIDO), and Design Space Optimization (DSO) that all have their place in CFD / CAE simulation workflows today; but they are all well over ten-year-old technologies and methodologies now. The diagram below shows the way modern AI/ML algorithms can be viewed in the context of this spectrum of DoE / PIDO tools and you can read more about the theory behind ODYSSEE in an AI/ML White Paper released by Kambiz last year. Indeed, ODYSSEE CAE includes all capabilities to the left of Machine Learning on the diagram as well as new and exciting capabilities and algorithms tailored for the CAE and manufacturing worlds. The recently released ODYSSEE A-Eye can fuse data from different sources such as virtual CAE simulation, real world scanning data, and even photos and video images; plus... financial and costing data can have ML and AI analysis applied in order to derive actionable outcomes very quickly! You can read more of how AI/ML has impacted metal forming simulation and costing in this recent FormingSuite magazine article by my Canadian colleague, Mike Lee.

ODYSSEE can actually measure the complexity of a model, a system or a process and uses a concept related to entropy in chemistry for robust optimization in ODYSSEE A-Eye of cost engineering for instance. Reducing complexity equates to increasing robustness in CAE models of any other processes. ODYSSEE software can even fuse multiphysics ROM (Reduced Order Model) data from different CAE solver types that historically typically would require very different timescales to converge due to the numerical intensity of their mathematical techniques – 3d FEA structures, 1d MBD, 3d FV CFD – in near real-time. This is a game changer in my judgement for the CFD and CAE industries by virtue of unleashing new levels of productivity without losing quality and opens the prospect of Operational Digital Twins that can democratize trained CFD and CAE simulations to non-experts such as plant operators. Hence, tens of millions of engineers and plant operators around the world, and even the general public, could benefit from the benefits of real-time CFD and CAE simulation!

What has astounded me the most in the last few months is the many customer examples I’ve now seen where 3d CFD simulation data (and other 3d point physics simulation types like structures, acoustics, crash, materials, multibody dynamics...) trained inside ODYSSEE CAE has led to interpolated and optimized predictions in seconds rather than minutes, hours and days that many of us are familiar with in traditional 3d CFD and CAE design space exploration approaches (and all do-able on a laptop!) This has involved 2, 3 and sometimes even higher orders of magnitude reductions of overall computational time to achieve these outputs. It has produced major productivity benefits for companies such as Yamaha in Japan and Stellantis in Europe – check out our recent Hexagon Design & Engineering Live 2021 conference AI/ML track for a myriad of real-world examples from companies such as Ford, Faurecia, Autoliv, Mahindra, Eta and assorted Universities. This approach has been proven in industries as diverse as automotive, aerospace, electronics, chemical process and the built environment.

Let me briefly outline two CFD related examples as illustrations, one from Europe and one from Japan, involving Cradle CFD from Hexagon, technically the most multiphysics-focused CFD code in the world, and ODYSSEE CAE:

Realtime Operational CFD Digital Twin of Hexagon’s Archidona Solar Farm

In this example, Hexagon created a 3d Cradle CFD model of our new Hexagon | R-evolution subsidiary’s Archidona Solar Farm in SW Spain - that was commissioned in September 2021 - by capturing Hexagon’s Leica BLK247 real world scanned data of the actual as-installed solar panel tracks on our 8 MWp solar farm. We took this real data into a Cradle CFD geometrical model of the plant and also incorporated 3d geo-scanning data for the neighboring landscape around Archidona to capture an accurate topological map of the hills around the solar farm site so we could fuse the real and virtual data. We then carried out a series of meteorologically bounded simulations for the solar plant that included the effects of wind direction, wind speed, solar radiation of the sun as it moves through the sky typical of different days across the year, and solar panels oriented at different angles to the vertical on a rotating single axis to represent their typical movement during the day at Spanish latitudes. Each full 3d Cradle CFD simulation took 1 hour to solve in the 3d Cradle CFD model on a large desktop machine but once a full CFD Digital Twin of the Solar Farm and panels was created inside ODYSSEE CAE, all sorts of “what if..?” questions could be determined for optimal operation of the solar plant on a laptop; typically only taking 2s per simulation.

This approach now allows us to couple real-world weather forecast data to next day solar farm operations in an actionable way to optimize electricity generation with our thermo-fluid Operational CFD Digital Twin of the site. It can therefore be directly linked to actions that maximize power generation efficiency on site, the profitability of the site, and ultimately, we have a powerful tool to educate Solar Farm operators onsite with real-time augmented reality CFD visualizations that they can hold up in front of themselves to understand what is happening in a thermo-fluid sense on or near to the panels they are viewing. This was done by coupling Hexagon’s powerful Xalt dashboarding technologies to Cradle CFD simulation predictions to provide a powerful AR visualization of the solar plant in real time on an iPad or indeed on an iPhone with the upcoming Cradle CFD 2022 product release this month. This fusing of real world as-installed and as-operated solar farm scanned data and our virtual Cradle CFD prediction data in a CFD Digital Twin is, we believe, a world first and a new Smart Digital Reality that gives Hexagon a unique means of optimizing the performance and outputs of our solar farm on a daily basis.

Optimizing HVAC duct location and flowrate inside a restaurant using CFD

In this second example, we were investigating where is the optimal location in a restaurant building to position HVAC air flow vents to ensure good mixing in the room – especially now in the era of COVID-19 - plus we wanted to determine the optimal duct vent flow rate. This is inherently a transient non-linear CFD simulation, and the complete 3d simulation took 40 minutes to solve on a 4-core desktop machine for one operating condition. A series of bounding 3d CFD simulations were then run and used to train ODYSSEE CAE for this built environment application. The 0d ODYSSEE ROM model interpolated the CFD predictions to provide almost the same level of fidelity as the 3d CFD simulations in just 30 seconds for other conditions and helped to determine the optimum air flow rate for this restaurant, table layout and vent scenario.

Concluding Remarks

If you want to know more about all of these technologies and have a CFD application that could benefit from being sped up several orders of magnitude, while improving your or your company’s CFD analyst productivity, then do call your local Hexagon | MSC Software office and ask more about the ODYSSEE suite of software and trial it yourself, or contact Kambiz or myself and see if we can do your CFD application in near real-time with AI/ML and Cradle CFD! The same applies for any of the Hexagon multiphysics simulation types mentioned in this article and also co-simulations between the different point physics types. All can be handled inside the ODYSSEE suite. I foresee a brave new world of CFD ahead of us all especially related to Smart Manufacturing that fuses real and virtual data in a seamless way and I believe that democratization of CFD is definitely within our grasp.