How will machine learning impact engineering simulation?
by Mike Worth, Senior Mechanical Engineer at 42T
In this article, we explore five ways machine learning will revolutionise engineering simulation. We shed light on how this new technology could be used to improve efficiency and get better results.
In recent years, machine learning has developed rapidly which has led to people exploring its utility in the field of engineering simulation. We explore what the prospects are for machine learning in revolutionising this branch of R&D and how it may impact the product development industry.
In many ways, the role of simulation in modern engineering is analogous to the slide rule or log table of generations past. While engineers use their judgement and experience to understand an engineering challenge and think of innovative ways to solve it, experience and intuition have their limits.
In most cases, a design is only truly understood by doing calculations, and in many safety-critical applications, calculations are a regulatory requirement. Simulation tools allow us to make our calculations far more advanced than those of our forebears, but this remains only one half of the story - successful engineering happens at the intersection of judgement and calculations.
Machine learning can do many amazing things, far surpassing what would be expected from 'just' pattern recognition. Its outputs can be scarily close to that of a human, and this brings us to the conclusion that it will never displace core simulation tools.
Creativity, estimation, and judgement can all be attributed to machine learning in many contexts, but these are all aspects of engineering carried out by human engineers - not their simulation tools. ML does not fundamentally understand calculations, and there are endless examples of AI chatbots confidently giving incorrect mathematical results. Yes, we are on a path to a machine learning empowered future, but one that still uses a traditional form of simulation solver.
We believe that the future of simulation is that ML takes on some of the human’s workload – it can streamline the inputs and outputs of calculations to improve speed, accuracy, and efficiency while reducing errors.
Creating and running a simulation is often a fiddly and repetitive task, but one that requires experience to get right. This is exactly the kind of problem that ML is good at - it can digest previous simulations, identify when particular configurations were applied and repeat that automatically.
The speed, accuracy, and ability to even get an answer are highly dependent on having a good set-up - the ML assistant won't save time by replacing the solver but will instead save time by configuring the solver so that it runs in a fraction of the time.
The top five
While there is a myriad of steps that simulation engineers complete, here are the top five places where I predict ML assistants will revolutionise the process:
1. Guessing initial conditions
Simulation calculations are typically iterative - the closer the initial conditions are to the solution, the faster and more stable the calculation is. Often, this initial guess is something completely false but easy to set up (e.g. nothing is moving). Using ML to get a close estimate to use as the initial condition will combine the speed of ML and the accuracy of traditional simulation.
The exact way in which a model is meshed is incredibly important, and the most efficient/stable meshes require an understanding of what the ultimate solution will look like. There are already tools that account for basic predictions such as typical fluid flows near walls and adaptive meshing that iterate meshes based on previous solutions. Using ML to guess at the solution upon which a mesh can be based could significantly speed up and stabilise the calculations.
3. Geometry clean-up
CAD files are typically created without detailed simulation considerations in mind. There are usually features included that actively hurt the simulation process, while not being important to the function of a real part. The challenge of geometry clean-up is that troublesome features are very situation dependant, and it requires experience to know which ones to adjust. ML powered 'auto cleanup tools' would save many tedious hours of manual work, while also substantially improving speed and stability of the calculations.
4. Identification of misconfigurations
None of us is infallible. Often the first run of a simulation fails because something needs adjusting. In some cases, this is a simple case of a typo or mis-click - in many other cases it's that one setting works better in a hard to predict way.
A machine learning assistant with a vast experience is likely to identify inappropriate configurations and provide warnings if things should be checked. Getting to a first result faster and more reliably empowers quicker iteration and agile working practices.
5. Unattended computation
This last opportunity takes a different approach to efficiency. There is a lot of spare processing power during down-time. Computers often sit idle overnight waiting for an engineer to look at the results in the morning and decide on the next thing to run. It is conceivable that ML assistants will take minimal inputs, and make decisions about what models to run, doing something useful with this 'dead' time.
Yes, there are already a plethora of optimisation and batch processing approaches that provide some of this value, but the overheads in setting them up dis-incentivise their use in many cases. ML with the ability to work with concise instructions and make sensible assumptions would make this far more routine and useful.
The simulation revolution
We are currently living through an exciting time for simulation - new ML approaches are emerging which will act to revolutionise the tool kits that we've all grown accustomed to in previous years.
In five or ten years, I'm sure we'll look back to now as the point at which things really started to change. We will have to adapt our role, but I see this as an opportunity for increased efficiency, improved results and a reduction in tedium.
Ultimately, no matter how much of our current workload is taken up by ML, we will always require human engineers to interface with the many stakeholders in what are ultimately human needs, and to be responsible for ensuring that designs are safe and appropriate.
If you would like to find out more, please contact 42T’s Robin Ferraby:
Robin leads 42T's work in the consumer sector, developing and managing our client relationships in the fast-moving consumer goods and appliance sectors.
Share this article: