Combining AI and simulation to automatically discover how processes can be further optimized
Can we speed up the process to get from simulation to optimization with AI?
Our client is a contractor at one of the US government’s most challenging nuclear waste clean-up projects. It’s responsible for the management of 56 million gallons of radioactive and chemical waste stored in tanks at the Hanford site, a 586-square-mile area in Washington State, dating back to World War II. The cleaning of this waste has started in 1989 and is estimated to take another 30-40 years. We are working for the engineering team whose mission it is to expedite the clean-up with the help of advanced analytics and simulation. In partnership with our colleagues from Lanner, we are deploying AI to speed up these analyses and help the team get better outcomes.
Working with the client team, we first mapped the process of getting from a simulation study to an optimization. Per process step, we investigated how we can make the process smarter or faster and how to approach that (it is always good to question the use of AI when there’s more simple methods that’ll do the trick).
“In a scrum process, three tools were developed that are used by our client’s analysts in their daily work.”
Connecting AI-models to simulations
First, we’ve developed a method to speed up the simulations themselves by replacing part of the model with an AI-based model. This not only shortens the run-time of a simulation, but also enables complex decision models to be included in the simulation.
On the long run, we’re aiming for this method to be used to train/test AI-models that will be used in real-time decision making in the production process.
Automatic optimisation discovery
Next, we’ve developed two solutions to automatically derive from a simulations outcome where there are opportunities for process optimization:
- Just like a real-life production process, a simulation can output real-time log data on what’s happening in the production process. That data forms the input to an algorithm that we’ve developed to spot bottlenecks in the system. For that, it for example looks at how machines are influencing each other negatively by keeping them waiting. The result of this analysis is an overview of which machines influence the rest of the process negatively most.
- Imagine training an AI model to predict the output of a production process, using a large amount of production parameters. Kind of like training a surrogate model to the simulation itself. When the model’s performance is good enough, you could research: how does the model know the output will be low today, but high tomorrow? By investigating what the models uses to make predictions, we gain insight into situations in the production that are related an optimal output and which situation are not. Likewise, for machine and process settings.
The results of these analyses are bundled into a tool that shows potential optimizations from multiple angles (“classic” algorithms and AI). What’s even better: this is not only applicable to simulation data but would work even better on real time production data, for example from a MES system.
Read more about these solutions and combining AI and simulation on our blog.
Our client faces a significant challenge to protect the natural environment by processing a large amount of nuclear waste. Data driven decision making is crucial to optimize and speed up this process. Our solutions aid in speeding up this data driven decision making. In the end to expedite the clean-up project, saving the environment and millions of dollars per day.