NEWSLETTER: February 2021

Hyperscale Computational Fluid Dynamics to Reduce Aircraft Emissions


As software names go, this slots straight in at number one. Sparselizard is one of the new kids on the block in the open source computer aided engineering software world, with a C++ based finite element library. The first line of code was written as recently as 2016, so it really is a new kid. With that in mind the capability list and user friendliness are very impressive, following a similar code abstraction approach to OpenFOAM, where you write the equation you want to solve in a succinct and easy to interpret format (if you know some calculus). The coupling capabilities across multiple physics domains is comprehensive, with a sizeable set of worked approaches available on the website and provided with the downloadable static library, covering elastic solids, contact, fluid dynamics, piezoelectrics, acoustics, thermodynamics, electromagnetics, optics and a variety of combinations thereof.


From a numerics perspective, sparselizard is "HP adaptive", allowing consideration of different polynomial orders (p-adaptivity) on elements that have been geometrically refined (h-adaptivity). An HP adaptive finite element method is a good thing as it possesses the ability to converge exponentially fast.  There's a decent development pipeline too, with scalability on many cores/nodes being targeted with the implementation of a data driven multi-threading (DDM) approach. In DDM, the computation and communication requirements of an algorithm are decoupled, allowing them to be executed asynchronously.


With solid foundations, usability and diversity of application, the future looks bright. A strong association exists between open source codes for computer aided engineering and OpenFOAM, it would benefit the open source community for some diversity here; sparselizard could well be part of that movement.

Monolith AI Partners with Siemens

There's a fair amount of excitement (you may alternately describe it as hype) around the use of artificial intelligence (AI) in a variety of applications. In the engineering domain there's a strong focus on capability being championed by Monolith AI, a London based start-up, who have been focused on engineering use of AI since their inception in 2018. They've recently announced a partnership with industry heavyweight Siemens Digital Industries Software, which will presumably result in some form of Monolith's web-based tool-set being made available through Siemens Simcenter portfolio offering, or a tight data integration between the two.


The partnership announcement was followed up by an example of Monolith's prediction capability, using computational fluid dynamics generated data for internal combustion engine design as input into Monolith's AI model. The methodology was sound, with the performance of the AI predictions assessed in a region of the design space on which the model was not trained. The output was impressive too with good agreement between the AI and CFD in this "test" region. The AI predictions consisted of key performance parameter relationships to the inputs, as they vary over time, and two dimensional field data of the flow characteristics in the piston chamber.


The results are impressive, but the level of accuracy begs an interesting question: how is confidence built in AI models that predict performance in the region of the design space where they have seen no data? Some of the approaches used for reduced order models (ROMs) could be followed, but maybe the expectation of AI models is that they embody the physics to a greater extent than ROMs; it will be interesting to see how this plays out.

A return flight from London to San Fransisco emits over five tonnes of carbon dioxide per person, over twice the amount produced by the average family car in a year. The emitted carbon dioxide is not the biggest contributor with regards to climate effect for aviation though; that comes from the formation of contrails, minute ice crystals formed as a result of the emitted water vapour rapidly saturating the moist ambient air and condensing and freezing as a result. These contrail clouds trap heat in the atmosphere, and account for around sixty percent of aviation's total climate impact, or around two percent of all human induced climate impact. Quite a good target to go after then.


Cambridge, UK, based Satavia are doing just that with their DECISIONX AI platform which uses weather prediction modelling to computationally replicate the behavior of the Earth's atmosphere over time. This weather prediction model is good old fashioned computational fluid dynamics, considering fluxes of heat, radiation and moisture and predicting temperature, pressure, velocity, humidity, aerosols and hydrometeors (cloud or rain to you and me). What's less old fashioned is the scale of the model; the grid spacing is around five kilometres on the planet's surface with sixty vertical levels going up to an altitude of one hundred kilometres, over the whole planet. That generates a nice four billion cell computational fluid dynamics model (run with a thirty second time-step, if the cell count wasn't impressive enough already.)


Unsurprisingly, Satavia opted for a cloud based capacity to solve such a large model, on Microsoft's Azure platform. Microsoft have form in this area; last year they ran a computational fluid dynamics model of hurricane Maria, comprising over three hundred and seventy million cells, on over eighty thousand cores, using AMD's second generation EPYC processors, with one hundred and twenty cores per processor. They used HDR Infiniband from Melanox to provide the parallel scalability required.


The computational fluid dynamics isn't the whole story for DECISIONX; the atmospheric model is used by an AI algorithm to provide two key capabilities. First, forecasting the expected contrail formation for given flight plans and recommending alternatives with a lower contrail footprint, to reduce the global warming contribution. Second, assessing the environmental exposure to aircraft for their flight plans and linking this to condition monitoring to enable more efficient maintenance schedules and engine operating conditions, both with a drive towards minimizing climate impact.


With a combination of deep physics modeling capability and AI model inference to recommend changeable flight plans, this looks to provide tangible climate benefits. The interesting part will be when the more climate friendly flight paths are longer than the originals.