Engineering Application of Artificial Intelligence & Machine Learning (Part-1)

Impact of Domain Expertise on Engineering Application of AI

Outline

Summary

Introduction

Part 1 — Solving Engineering Related Problems Using AI & Machine Learning

Data-Driven Modeling of Physics

Smart Proxy Modeling

Hybrid Models; Bad Idea

Summary

Since Artificial Intelligence is defined as the simulation of human intelligence, mimicking “Human Brain” for Analysis, Modeling, and Decision Making, it becomes important to note how human intelligence is used to solve engineering related problems. The main question that will be addressed in this article is the differences on how Artificial Intelligence needs to be used to solve engineering versus non-engineering related problems. If there are logical differences in how human intelligence is used to solve engineering and non-engineering related problems, then it makes sense that the same logic be applicable to Artificial Intelligence. This article also covers the requirements of Engineering application of Artificial Intelligence and its current problems in academia and industry.

Introduction

Let’s start by defining “Artificial Intelligence” and “Machine Learning”. “Artificial Intelligence” is the simulation of human intelligence mimicking “Human Brain” for Analysis, Modeling, and Decision Making. “Machine Learning” is the science of getting computers to act without being explicitly programmed through using Open Computer Algorithms to learn from data instead of explicit programming.

What today is called Artificial Intelligence and Machine Learning started as concepts and ideas in the early 1950s. The rule based “Expert Systems” ended up being referred to as Artificial Intelligence while “Perceptron” (data-driven pattern recognition) was moved out of this technology due to a book written by Marvin Minsky and Seymour Papert in 1969. Some research continued to be done on data-driven pattern recognition and finally this technology addressed the issues identified by Minsky and Papert and started to be what it is today, in the mid-1980s.

In the early to late 2000 AI and Machine Learning was used extensively in some games (Chess, Go, …) and the internet (Google, Image Recognition, …). In the past couple of decades, the overwhelming majority of applications of Artificial Intelligence and Machine Learning that the society has been exposed to, has been used to address non-engineering related problems. The engineering application of Artificial Intelligence and Machine Learning is quite different from how this technology is used to solve and address non-engineering related problems that is also known as human-level intelligence, or Artificial General Intelligence (AGI). Engineering domain expertise is an absolute requirement in the engineering application of Artificial Intelligence and Machine Learning.

Unfortunately, recently it has been learned that domain expertise can have both positive and negative impact on engineering application of Artificial Intelligence and Machine Learning. This will also be discussed in this article. Domain expertise impacts engineering application of Artificial Intelligence and Machine Learning positively when the engineers develop a solid understanding of the philosophy, major characteristics, and realistic approach of Artificial Intelligence and Machine Learning. On the other hand, domain expertise has been impacting this technology negatively, when the engineers come to the conclusion that only traditional statistical knowledge and understanding of the mathematics behind the Machine Learning algorithms is all that is needed to make them an expert in the AI&ML. Furthermore, many traditional engineers refer to AI and Machine Learning as a new tool that needs a bit of skill, and they believe it is a tool like what Excel used to be and it’s easy for everyone to learn how to use it.

Application of Artificial Intelligence and Machine Learning in Engineering related problems is not confined to applied statistics and mathematics. To maximize efficiency and spur practical innovations, engineering domain experts must be seriously trained to become expert Data Science practitioners. Objectives of Engineering Application of Artificial Intelligence & Machine Learning include:

● Modeling physical phenomena using facts, reality, and measurements in order to avoid any assumptions, interpretations, preconceived notions, and biases.

● Advancing the art and science of engineering problem solving, design, and uncertainty quantification with extensive incorporation of Machine Learning algorithms.

● Training the next generation of engineers and scientists with practical knowledge and expertise in the art and science of Artificial Intelligence & Machine Learning.

In this article solving engineering problems using AI and Machine Learning is explained, followed by identifying what are the key requirements of this technology to solve engineering related problems. Furthermore, the reasons behind the lack of success of this technology in some of the industries will be discussed.

Solving Engineering Related Problems using AI & Machine Learning

Today’s traditional approach of engineering problem solving that was initiated during the sixteenth century, has significantly contributed to the industrial revolution throughout the mid eighteenth to nineteenth centuries. Engineering problem solving requires modeling physics. The major approach to modeling physics includes identification of the variables (parameters) that control, influence, and impact the physical phenomena and then mathematical equations are used to model relations, interactions, and influences of the variables on one another. Complexity of the physics that is being modeled determines the complexity of the mathematical equations (linear equations, non-linear equations, ordinary or partial differential equations, etc.) that are used to model them.

As researchers and scientists develop better and more realistic understanding of the physical phenomena that is being modeled, it usually helps to modify the mathematical equations in a manner to provide better, more realistic and useful solutions. Scientific research enhances our understanding of physics as a function of time (scientific understanding of physics enhances as time moves forward). Artificial Intelligence and Machine Learning generate a completely new approach to modeling the physical phenomena that is quite different from the traditional approach to engineering problem solving. The correct way of using Artificial Intelligence and Machine Learning to solve engineering related problems completely avoids the development, generation, and use of mathematical equations for modeling physics.

● Data-Driven Modeling of Physics; This technology uses actual measurements in order to build a model, avoids mathematical equations

When Artificial Intelligence and Machine Learning is used to generate predictive models for solving engineering related problems, using mathematical equations are completely avoided. Furthermore, unlike the traditional approach to engineering problem solving, the variables and parameters that are used during the AI and Machine Learning engineering problem solving are not only limited to explicit variables and parameters that directly contribute to the solutions. In this technology even implicit variables and parameters can be used during the Predictive Data Analytics.

Using Artificial Intelligence and Machine Learning to solve engineering related problems is purely based on “Data”. This is the main reason that this technology is called “Data-Driven”. On the other hand, the main characteristic of the data that is used in this technology is “Actual”, “Measured”, and “Collected” data rather than data that might be generated using mathematical equations. One of the questions that is commonly asked is:

“Why shouldn’t we use data generated by mathematical equations in order to develop data-driven solutions for a given problem?”

This question can be answered in multiple ways. First, and foremost would be:

“You do not need to develop a Data-Driven model and solution, if you do not have enough real data.” Furthermore, one can say, “if you believe 100% that the mathematical equations that solve the physical phenomena are correct, then why would you want to use Data-Driven AI-based modeling? unless you are using it as an academic approach to teach students how a Data-Driven approach works.”

Second answer is:

“When you include certain amount of data that is generated by mathematical equations, then you are making biases on how the solution should be accomplished, since data generated by mathematical equations, pushes the Machine Learning algorithms in specific and predetermined directions.”

And finally, the answer is:

“The main objective of the AI-Based modeling of physics is to avoid assumptions, interpretations, simplifications, preconceived notions, and biases. When you use mathematical equations to generate data and then include such data in the AI-base modeling process, you are undoing the main reason behind using Artificial Intelligence and Machine Learning to solve engineering related problems.”

The fact is that the overwhelming majority of the “Actual”, “Measured”, and “Collected” data from a given physical phenomenon, usually include the characteristics of the physical phenomenon that is going to be modeled by Artificial Intelligence and Machine Learning. If you do not have enough “Actual”, “Measured”, and “Collected” data, then using mathematically generated data to have enough data to use this technology would only have academic use. However, it should not be used or claimed to have actual and realistic use for solving engineering related problems using Artificial Intelligence.

● Smart Proxy Modeling; This technology uses data from mathematical equations (analytical or numerical solutions) in order to replicate the traditional models.

Smart Proxy Modeling is the application of Artificial Intelligence and Machine Learning in numerical simulation models. Smart Proxy Modeling is a purely data-driven modeling technology with one major difference with what was mentioned in the last section. Instead of “Actual”, “Measured”, and “Collected” data, Smart Proxy Modeling uses the data that is used and generated by Numerical Simulation. Smart Proxy Models generate the same solutions that are generated by the numerical simulation models with very high accuracy and with orders of magnitude less time (minutes on your workstation and laptop instead of multiple hours on HPC).

Historically, lack of speed for generating results using numerical simulation models was the reason behind development of several traditional proxy modeling techniques. Two of the mostly used traditional proxy models are known as Reduced Order Models (ROM) and Response Surface Method (RSM). Development of Reduced Order Models usually changes the physics of the numerical simulation model through simplification of its mathematical equations and /or reduces its resolution in space and time. While Reduced Order Models (ROM) makes a numerical simulation model to be deployed much faster, the results it generates are usually quite different from the actual results of the complex physics and high resolution of the numerical simulation model.

The second common proxy model is Response Surface Method (RSM) that is a traditional statistical technology. The problem with Response Surface Method (RSM) is that it concentrates on a specific part of the numerical simulation model in space and time rather than proxy modeling the numerical simulation model in its entirety. Furthermore, the other main problem with Response Surface Method (RSM) is that it requires hundreds of runs of the numerical simulation model for its generation, and even then, the results it generates leaves much to be desired.

The key behind Smart Proxy Modeling that distinguishes it from the traditional proxy modeling (ROM & RSM) is the fact that (a) it does not change the physics or the resolution (in space and time) of the numerical simulation model, (b) it replicate the entire numerical simulation model in space and time, ? its accuracy is orders of magnitude better than the traditional proxy modeling, and finally (d) it can run (deploy) a highly complex numerical simulation model in few minutes on a laptop or a desktop workstation.

● “Hybrid Models” combining AI with Data Generated by Mathematical Equations. Bad Idea.

When a traditional engineering approach requires mathematical equations to model physics and the AI-based modeling of the physics is purely based on measurements (hard data) it makes no sense to combine these two techniques since it would completely negate the main reason behind using AI to solve engineering related problems. The main reason behind using AI & Machine Learning to model physics is to avoid any possible assumptions, interpretations, and simplifications that are part of the main characteristics of traditional modeling of physics. This does not mean that traditional modeling of physics is a bad idea and must be avoided. The idea is that when a new technology has been developed in order to avoid certain characteristics of the previous approach, then it makes no sense to bring in and include the same characteristics that this new technology tries to avoid.

The main characteristics of AI & Machine Learning that is used to model physics has to do with this technology’s ability to discover hard and complex patterns and trends that might be part of the measured data that is used to model the physical phenomena. Given the fact that traditional modeling of physics that uses mathematical equations provides relationships (trends and patterns) between all the identified involved parameters, then when these mathematical equations are used to generate data to be combined with actual measurements, the predetermined and biased patterns and trends (developed by the mathematical equations) are being combined with the actual measurements and will guide the process to a certain direction. This means that the person or the company that uses “Hybrid Models” is avoiding the main characteristics of modeling physics through AI & Machine Learning.

It would be interesting to ask a simple question from those that try to combine data generated through mathematical equations with actual measured data to solve engineering related (physics-based) problems. Question: “if you would have been able to solve the problem only with actual measured data, would you still use hybrid-models through combining the actual measured data with data generated through mathematical equations?” It seems the answer to this question should be quite obvious. If they are capable of solving the problem using actual measured data, why would they do anything else? Therefore, it seems obvious that the reason behind developing “hybrid-models” is the lack of success of using Artificial Intelligence and Machine Learning to solve engineering related (physics-based) problems.

There seems to be two possible reasons behind the development of “hybrid-models”. First and foremost, it seems that those that have done this, apparently have a very limited understanding of AI & Machine Learning’s characteristics in general and definitely, have very little to no understanding of characteristics of this technology when it is used to solve engineering related problems. On the other hand, if they (those that use “hybrid-models”) have a good understanding of the science of AI & Machine Learning, but still use “hybrid-models”, then it determines that they are using the term “AI & Machine Learning” only as a marketing tool rather than from a scientific point.