Techniques like PID control or MPC can track single or multiple setpoints effectively. However, it remains a great challenge to adaptively find the optimal setpoints, or in batch process context, the optimal recipe, for these lower-level controllers to track.
Either coming up with a good recipe from engineering runs or starting with that recipe and then fine-tuning at production scale is essentially an optimization problem. Usually, we would mostly just rely on the domain knowledge to generate the initial recipe, do some small-scale lab runs, and then just run with the best setpoints for scaled up production. There are a few problems with that approach. Firstly, there are unavoidably differences between engineering runs and production runs. Secondly, some process noises would change with time. Factors like aging of the equipment, raw material suppliers or maintenance may cause the “optimal setpoints” to slightly shift from time to time.
Can “Modelling with Big Data” solve this problem?
Can we approach the problem by collecting all kinds of data possible on the recipe-controlled production runs and then leave it to the Big Data/Data Science team to come up with the best solution? Well, not really, even with an infinite amount of data. Because if all the data you can collect are controlled by the same recipe, you will probably have a great model predicting what happens for the setpoints based on that recipe but won’t have much luck in predicting anything else when the setpoints are different than that recipe. In other words, your machine learning model will have a hard time extrapolating. Hence, it is not only the data volume that matters, but the variability matters even more here for optimization.
What may be a good approach to this problem?
Mostly optimization problems don’t care as much about the journey to the optimal solution as the final optimal solution. For example, if the problem requires 100 optimization iterations, the algorithm may not care about the performance of the first 99 iterations if the 100th iteration yields a good result. It is not the case here since we care about the performance of every iteration (batch). In other words, it is the average performance of all batches that we want to optimize, not just the max performance. For scenarios where good mechanistic models (digital twins) are available, the problem does become much easier since you can potentially do most of the experiments on the mechanistic model before running on production. However, building a good digital twin representation is a great challenge by itself.
Despite all the challenges, we believe that better outcomes can be achieved when combining recent advances in AI and Machine Learning. And we categorize the approaches into 3 levels depending on how mature the plant/process is digitally.
Level 1. Only the outcome and setpoints are recorded.
As hard as it is to believe, some plants don’t have robust and consistent measurements from sensors or other data sources that can be used for optimal control. Good news is that there are still techniques that if designed and applied properly, a good performance boost might be achieved. At Quartic AI, we tackle this problem with Bayesian Optimization. Bayesian Optimization is one of the most effective choices when it comes to optimization where the dynamics of the process is not observable. It of course comes with its caveats and we will discuss how we address them in our next blog post.
Level 2. Some or most of the key measurements are available (+ Level 1).
When you have good quality of key measurements, you can start thinking about more interesting approaches for optimization. You can build on top of your Level 1 approach (e.g. adding context to Bayesian Optimization), or design hybrid modeling approach with Machine Learning, Reinforcement Learning and MPC. We will publish our research and case studies at this level in future blog posts.
Level 3. High-Fidelity Digital Twin (+ Level 2).
For some processes such as bioprocessing, it is quite challenging to have a high-fidelity digital twin to represent the entire process well enough to be useful for optimization. However, if we have it, it will potentially make productizing the cutting-edge research in Machine Learning and Reinforcement Learning a reality. It also makes things like Root Cause Analysis much easier which also contributes to optimization in a broad way. Also stay tuned to our research and findings at this level.