Recently, our mother company Royal HaskoningDHV acquired the UK-based predictive simulation expert Lanner. Whereas we at Ynformed are focusing on applying AI/machine learning (ML) techniques in various domains, Lanner develops the well-known simulation software WITNESS and helps clients globally to optimize their processes using simulation.
For us at Ynformed this is a really interesting development (and I’m assuming for Lanner too ). Although we are obviously big advocates of applying AI to solve problems, there is a whole branch of problems that are better solved using simulation. E.g. when there just isn’t any data to train AI-based models on, or when the dynamics of a system are just too complex to fit a model to. Also, when you want to model for optimization, we sometimes see people misemploy machine learned models where simulation should have been used*.
* Suppose you want to optimize the parameters of a wastewater treatment facility for optimal output values. One approach could be to train an ML model using the systems inputs (e.g. inflow volume), parameter settings (e.g. pump speed) and outputs (e.g. nitrogen level). It is tempting to use the resulting model for optimization: if we tweak the model’s inputs (e.g. double the pump speed), what happens to the output? While for simple systems this might yield realistic outcomes, in more complex systems this might go very wrong. Especially if the input values are outside the range of what a model was trained on, which typically is the case for optimizations, a model might extrapolate incorrectly. In short: the model has never seen a double pump speed, so it may be completely wrong with the corresponding outputs, without the users noticing.
Combining AI and simulation
But what really got us thinking was: can we combine simulation and AI? What if we for example could use simulations to train our models(as is happening in reinforcement learning), or the other way around: can we see our models in action in simulated environments before deploying them?
Together with Lanner we came up with about 8 possible ways of combining AI and simulation. For now, I will go deeper into two of those possibilities. For a US-based client we are currently investigating two combinations of AI and simulation:
- Can we speed up the optimization process by having an AI analyze simulation outcomes and point out possible improvements?
- Can we speed up/extend simulations by replacing time-consuming models with fast AI approximations?
Speeding up the optimization process
The client has already developed simulation models of their facilities in great detail. Amongst other things, these are used to analyze possible optimizations to the facility. Now imagine being an analyst and having to look for possible optimizations in 10.000 simulated years of facility operations, with hundreds of machines interacting. Quite hard already. But you’d want to do this continuously, to iteratively come up with improvements that you can test. That is really hard and time consuming!
To deal with this, we’re researching the application of AI to aid this process. Models that can automatically identify areas for improvement when going through the simulation outcomes. For now, with the goal of making the analyst more effective. But imagine if we could generate new scenarios based on these identifications and feed them back into the simulation. We would effectively have an AI-based optimization engine, applicable not only in industry, but in many other domains too.
Extending simulations with fast AI approximations
The image above is probably familiar. Missing in this image, but typically also an aspect, is that a machine learning model can be much faster than a computed “traditional” model. That might come at the cost of lower accuracy, but in many cases that is not a problem: going from a 15 minute to sub-second compute time is worth it. Especially if we’re considering a model that needs to be called for decision making inside the simulation numerous times: your simulation would simply take too much time to run!
We’ve encountered this situation not only in industry, but also in the other domains we’re active in. Whilst there’s no generic solution to this problem, and we might not be able to improve the situation for really complex models (as explained before), we expect that we can make good progress on this subject.
AI and simulation become parents to twins, digital twins!
It is not hard to see how these AI infused simulation models, fed with real time data, provide prescriptive functionality to digital twins of physical facilities and assets. An example situation:
Twin: “I’ve detected an oncoming failure for pump A in the facility: it will fail in 90 days” (AI)
Twin: “We’ve also detected this is a critical part of your process, you might want to look into replacing it with a bigger pump” (AI + Simulation)
Twin: “Scenario runs show that an optimal time for replacement is in 60-65 days from now” (Simulation)
Let’s stay in touch
We feel like we’re just scratching the surface of what’s possible. As we progress with first clients, we will keep you posted on what is happening. If you take a special interest into these topics or want to see what we can do for you, don’t hesitate to reach out to me at firstname.lastname@example.org.