I see machine learning (ML) as the last and most complex resource when developing a process model used on a specific application (for instance, process control) to estimate a process variable value (also known as soft sensors) or to predict failures. The key point is that there should be a proportionality criterion between the engineering solution and the problem to resolve. While overdesigning takes more effort, time, and money, nonetheless the achieved benefits are not in the same order of magnitude as the engineering labor. Depending on process complexity and model performance requirements, modeling techniques vary from simple to elaborate.
First, let's define a complex process/phenomenon: A system that heavily depends on initial conditions, is non-linear, has some behaviors that are stochastic, and has the sum of its parts bigger than the total.
For well-known deterministic industry processes, models can be obtained by applying first principles. However, due to model uncertainty, or modeling simplification (e.g., linearization, approximations) the capability to emulate the process, as per required performance, is negatively affected. To compensate this gap, the model can be completed by adding a function attained from data regression and/or a probability distribution function.
When the aforementioned attempts are not capable to produce a suitable model per application requirements, a more sophisticated method is necessary. Here is where ML is an option to consider. Of course, it has to be justified. In other words, it has to be feasible to develop a model and the potential benefits need to be reasonably achievable.
ML is a useful model built from data. There is a dataset to train the model and a separate dataset to test the model. When the model is deployed, it is intended to replicate the process, even if the inputs values are different from those used to train the model. In other words, it is expected to generalize with a pre-established performance accuracy. As ML models can be affected by prediction errors or overfitting, model parameters can be adjusted to improve its performance.
As mentioned, if with conventional techniques (i.e., first principles, linear regression, or statistical models) a suitable model cannot be obtained to fulfill application requirements, then ML is an option that can be assessed.
ML could be seen as a black box, as its implementation is in the form of programing code (e.g., Jupyter Notebook) or connecting diagrams (e.g., Microsoft Azure). If one reads the coding, they will not be able to directly read the model structure (input and output variables, how they are correlated, the constraints, etc.), whereas in conventional modeling and programming, model setting can be understood with minor effort.
Here are some examples of potential industrial applications for ML:
Engineered solution complexity increases with problem complexity, ideally in a linear proportion. Albeit it might be exponential for some particular cases. ML could be overkill for applications where conventional techniques are suitable, as ML additional effort can only deliver the same benefits as conventional techniques. Any applications has to be assessed on a case-by-case basis.
*Note: These instruments can also be developed based on first principles, therefore achievability and benefits will dictate where to go.