Machine Learning is not to compete with a formal mathematical equation.

        By Pedro Perdomo

 When you want to have a digital model of your process, mainly you have two options, real three but let’s take the first two.

You can use the first principles of physics and try to get a fairly, replicable, and “precise” math model. On the other hand, you can use data to train a machine-learning algorithm and get a math actuation,  a black box,  but certainly wit an embedded rich acknowledge of your process. The first option needs subject matter experts who understand the thermodynamics of the process. The second just need data experts.  Which option to take? Why?

In engineering, mathematics does not have all the solutions. On many occasions, we understimete variables in an math equation looking for a practical solution, close to reality. We do NOT struggle to calculate milligrams if at the end, we can NOT measure milligrams.

Making a mathematical model of an industrial process suggests posing primary equations, well-understood thermodynamic relationships of chemical and physical processes that describe the properties of fluids under certain conditions of pressure, flow, and temperature. In many of these equations, factors and relationships of difficult determination are used, many come from measurement and practice, and suffer from uncertainties and boundary conditions. 

 

Experts in the mater are needed to develop them. Even so, they are not infallible of errors.

 

 Here is the beauty of Machine Learning. Instead of those formal mathematical equations based on thermodynamics principles developed by experts, we only need DATA, enough and “rich data”.

Offshore Platform Twin

 

 Machine learning algorithms describe the world in some hidden way, not entirely understandable, but in most cases even more accurate than mathematical models developed by human matter experts.

 

 At the end of the day, the outcome of machine learning is a mathematical equation, simple, but thermodynamically incomprehensible, black boxes very difficult to interpret it, but certainly, wit an embedded acknowledge of the process.

 

However, many mathematical relationships within the process arena are already very well known, with well-defined basic first principles equations that are easy to implement (white boxes). Some of the "factors" and uncertanties can be dismissed or estimated and the result yet of great value. Pumps, furnaces, heat exchangers, coolers, are some of them. I am not trying to make a rigorous contrast against formal thermodynamic equations yet take dvantage of new technologies.

So which way to go? Thermodynamic or data experts.

I leave you some considerations, some  pros and cons:

A model with machine learning (data-driven)

PROs

• Model taking from data only - we do not need Newtonian equations

• Generic and flexible - handles heterogeneous (also non-physical) data flows

• The model improves over time (learning reinforcement)

• Good for discovering complex relationships and patterns

CONs

• The availability of training data to develop the model

• Correlations, not causality. Blackbox, no explanations (in particular, deep learning)

• Approach methods, no exact math

• Predictive capabilities deteriorate rapidly outside the scope of the training set

• Difficult to predict extreme / critical conditions (few observations)

• Biases in data is reflected in the prediction

• Poor generalization of unseen problems

Physics-based model

PROs

• Los models capture existing insight based on Newtonian physics

• Causal relationships provide insight and understanding

• Input-controlled uncertainty and modeling accuracy

• Model has universal validity: predicts any point covered by the model

• Generalizes well to new problems with similar physics

CONs

• Requires extensive knowledge Mastery (physical)

• Computationally intensive, a real-time challenge

• Full I / O assumptions must be made in advance

So, what is the best way to follow up?

It depends on many factors, inclusive human factors not mentioned above. Good engineering criteria is necesary.

But why do not take any combination of these alternatives?

I do think it will be the best approach. 

If you are taking the chance to jump into this wave to help you up in having an express math model of your process using just data, it is because you are decided to make the most of your assets, you need to know them deeply.  You are taking advantage of a tool, you know your process, your assets. If you do not, much probably you won’t be able to use the data the model is given back to you.

Do not replace the acknowledge of your process and industrial assets by just data models, just take advantage of it.