{{ post.title }} 글 편집 Author {{ post.author.name }} Posted on

CAE Simulation and AI

What are some terms that are related to CAE? While the precise moment may not be knowable, at some point the terms that come to mind in relation to CAE began to take on a different nature. While such as finite element method, multi-body dynamics, fatigue analysis, computational thermo-fluid analysis, optimal design, and automation were associated with CAE in the past, in the present day terms such as artificial intelligence, machine learning, data science, Internet of Things, and surrogate models are now associated with CAE.

When these terms are grouped together, differences emerge. One group can be labeled “CAE terms,” while the other group can be labeled “4th Industrial Revolution terms.”

So what sort of relationship do the 4th Industrial Revolution terms have with CAE? In order to answer this question, these terms have been defined as follows. Hopefully, the information presented here gives those involved with CAE the opportunity to re-examine how their responsibilities relate to these terms. And learning about these terms can be helpful even for those who are not involved with CAE, as these terms have recently been the focus of much media coverage.

1. Simulation

A Google search for the term “simulation” returns, “Expressing a physical or abstract system as a model, and using this model to conduct tests. Simulations include physical simulations done through the creation of physical models and theoretical simulations done through mathematical models on computers. There are also real-time simulations used for engineering design and social phenomenon analysis where massive numerical calculations are done at extremely high speeds.”

Such simulations are essentially “virtual tests.” Among such virtual tests, virtual tests that utilize computers are called computer simulations . And in particular, the use of computer simulations to design and analyze products is called Computer Aided Engineering ( CAE ). In other words, CAE can be considered as a type of computer simulation.

(Someone calls CAE, CAE Simulation or CAE Simulations)

2. Digital Twin

Digital Twin is a concept created by the US company GE, and means the creation of a “twin” within a computer that is identical to its real-world counterpart, and the use of this twin to conduct computer simulations of situations that can occur in the real world, thereby predicting the outcomes of such situations in advance.

The LG CNS blog cites materials from Gartner in showing the levels of execution of the digital twin concept, with the levels being: Level 1, 3D visualization; Level 2, real-time monitoring; and Level 3, analysis/prediction/optimization.

<Level of Digital Twin Realization > ( https://blog.lgcns.com/1864 )

Source: Gartner, Use the IoT Platform Reference Model to Plan Your IoT business solutions, September 17, 2016)

The difference between Level 1 and Level 2 is whether or not online methods are used to collect the data that is to be input into the model. Using offline data to conduct simulations in advance , thereby allowing for 3D visualization, is Level 1. Level 2 involves models applying online data obtained from sensors on actual objects through the use of IoT platforms. At this level, the object and the model are both subject to the same experiences, so the actual object and the digital twin can be seen as a 1:1 match. Level 3 involves the use of input data and results to predict results in the future.

3. Cyber Physical System (CPS)

While not as well-known as Digital Twin, there is also the concept of the “Cyber Physical System.” CPS means an integrated system where computers use digital communications to control physical systems. To be more precise, CPS means a system that controls or monitors a mechanism through the use of computer algorithms. Software also works in concert with actual physical systems in CPS as well.

But CPS seems similar to Digital Twin, so what are their differences? Forum materials from KCERN on Digital Twin and Smart Transform explain the differences between Digital Twin and CPS as follows.

• Using data collected during the digitization step to establish a virtual reality that is 1:1 with reality is called Digital Twin.
• While Digital Twin establishes a virtual reality, CPS goes beyond Digital Twin to connect virtual reality with reality.
• CPS is the process of realizing in reality the optimized figures that were obtained in virtual reality by the Digital Twin. During this process, artificial intelligence is used to analyze data while other technologies are used to realize the virtual reality in actual reality.

Based on this explanation, it appears that, “Digital Twin is the virtualization of reality, while CPS is the realization of virtual reality.”

4. Meta Model

CPS, which was explained above, is called a system of systems. In contrast, the meta models and surrogate models addressed in this section are called a model of a model. While a system of systems ties systems together, thereby increasing the complexity of the system overall, a model of a model means simplifying a model. Approximation methods are used to simplify models. So meta models are also called approximate models . In other words, an approximate model means modeling a model once again in order to simplify it.

Approximate models are often used for optimization. If the number of repeated calculations utilizing the real-world model is increased during the process of optimization, then a much longer amount of time is required, so an approximate model is used to reduce the time needed for calculations.

The models used in the CAE simulation are meant to provide accurate results. So conducting detailed modeling without distortions requires a large amount of computation and, in turn, a large amount of computing time, so this is a computation-intensive method. If optimization is done through computation-intensive methods, the amount of time required can increase to multiple months, meaning that such time intensive methods may be difficult to utilize in actual business.

So as an alternative, optimization can utilize approximate models instead of the computation-intensive models. By using approximate models, the amount of time needed for optimization can be greatly reduced. The approximate models that are used for optimization in particular are called meta models.

But regardless of whichever simulation model or method was used, only the system’s response is used to create a meta model. Inputs and responses are used to create a meta model, and this is used to conduct optimization. This method is called approximation based optimization.

examples)

• RecurDyn/DriveTrain that uses Meta Model for gear contact analysis
• RecurDyn/AutoDesign uses Meta Model for optimization

5. Surrogate Model

Surrogate models, along with meta models and response surface models, are another type of approximate model. As with meta models, approximate models take the original model and model it once again.

Surrogate models are formed through data driven, bottom-up methods. They are unconcerned with the internal logic of the simulation and simply focus on the inputs and outputs. A model is formed by using the results associated with the engineering variables. Due to these characteristics, this is also called behavioral modeling or black-box modeling. And cases involving a single engineering variable are called curve fitting.

Meta models and surrogate models have the same definition in a conceptual sense. But while their definitions are identical, meta models are mostly used for optimization while surrogate models are used in a more general sense in the engineering field. So in fields outside of optimization, the surrogate model is the better known term.

6. Deep Learning

In his book “Deep Learning for Everyone,” Professor Taeho Jo of Indiana University gives the following explanation of artificial intelligence, machine learning, and deep learning.

Deep learning is the result of decades of continuous work to create artificial intelligence that resembles humans. While researching artificial intelligence that would allow computers to make decisions similar to those humans can make, researchers discovered that it was effective to utilize machine learning methods that used existing data to predict future events. There are many algorithms for such machine learning, but the most effective among them is deep learning. Therefore, the relationship between artificial intelligence, machine learning, and deep learning can be expressed in the following manner. Machine learning lies within the scope of artificial intelligence, and deep learning is a part of machine learning.

Terms like artificial intelligence, machine learning, and deep learning are being used interchangeably, and it is convenient to conceptualize them as follows.

<Relationship between AI, Machine Learning, and Deep Learning>

7. Machine Learning

Machine learning is a subsection of artificial intelligence, and is the collection of algorithms and technologies that allow computers to learn. For example, machine learning allows users to train the computer to recognize whether or not a received email is spam, and allows for the development of algorithms that can detect defects by training over multiple sets of simulation results.

Machine learning and data mining use the same methods. Computer science uses the term machine learning, while statistical science uses the term data mining. If machine learning and data mining were to be distinguished, data mining aims to discover forms and characteristics within the existing data, while machine learning uses existing data to train in order to identify expected values with respect to new data (Euijoong Kim, “Introduction to Artificial Intelligence, Machine Learning, and Deep Learning” ).

8. Artificial Intelligence (AI)

From the perspective of professionals, artificial intelligence has many definitions. In their book “Artificial Intelligence: A Modern Approach,” Stuart Russell and Peter Norvig classify artificial intelligence into the following 4 types.

1) Systems that think like humans

2) Systems that act like humans

3) Systems that think rationally

4) Systems that act rationally

Among these, the most active area of artificial intelligence research is with “systems that act like humans.” For example, this includes natural language processing, image recognition, voice recognition, machine translation, computer vision, robotics, etc.

Artificial intelligence is already in our lives. From the smartphone feature that can identify flowers by just taking a photo, to natural language voice activated digital assistant systems such as the iPhone’s Siri, all of these solutions utilize artificial intelligence.

But more than anything else, the reason why artificial intelligence is a widely known term among people is because of Google’s artificial intelligence Alpha Go and its Go matches against humans. By prevailing over humans in these competitions, people began to feel threatened by artificial intelligence and society responded to these feelings.

9. Prognostics and Health Management (PHM)

PHM is a technology that collects state information from machines, equipment, airplanes, power plants, etc., to detect irregularities within the system and predict points of failure in advance through analysis and advance diagnosis in order to optimize facility management (Korean Society for Prognostics and Health Management website http://www.phm.or.kr/ )

PHM monitors a machine’s status in real-time, allowing for advance detection of irregularities such as vibrations or wear, and can predict failure that may occur in the future. If failures can be predicted, then appropriate measures can be taken in advance, so unnecessary repair and maintenance costs can be avoided and reliability can be enhanced.

And just like with Digital Twin and CPS, sensors and IoT are important for real-time monitoring in PHM as well. And among the types of artificial intelligence, machine learning in particular can be used for predictions.

Research into PHM was started by the UK’s Civil Aviation Authority (CAA) in an attempt to reduce the accident rate for helicopters, which was 30 times the accident rate for aircraft at the time. The CAA developed HUMS (Health & Usage Monitoring System) to monitor the soundness of helicopters, and after using this system on actual helicopters, found that the system reduced the accident rate by 50%.

The foregoing was an examination of these 9 terms; the following is a visual representation of what was examined.

<CAE simulation and AI>

• Simulations bring a real world subject into a virtual world in order to conduct virtual experiments.
• Bringing a real world subject into a virtual world is the creation of a model.
• As such, the model separates the real world from the virtual world.
• The virtual world as seen from the viewpoint of CAE is the Digital Twin.
• Digital Twin has 3 levels.
• Currently, most CAE receive their inputs offline, so they can be viewed as level 1.
• Level 2 involves the use of the same computation-intensive CAE model as is used in level 1, but in this level, the inputs are received online through IoT.
• Level 3 involves making predictions from the modeling results.
• Predictions occur through the use of machine learning. Machine learning is one sector of artificial intelligence.
• Just like in a surrogate model, machine learning involves analyzing the relationship between the given input and output data to make predictions.
• In order to continue machine learning through training, new data must be provided.
• CAE simulation can be used to provide the data sets to be used to train the machine learning.
• A Cyber Physical System (CPS) takes Digital Twin predictions and applies them to the real world.
• For example, monitoring irregularities with a machine or plant and making failure predictions (PHM) is the responsibility of the Digital Twin, while the CPS uses these predictions as a basis for reducing speeds or suspending operations in order to prevent a failure from occurring.

In light of the above, CAE will most likely increase in importance in the future, as it can prevent accidents and failures from occurring while ensuring user and machine safety. In the future, CAE will most likely be included as a part of artificial intelligence. As rapid response times are required when implementing predictions through artificial intelligence, there is a high likelihood that the use of models of a model, or in other words, surrogate models, will increase in the future. If so, then in order to increase the accuracy of the predictions, machine learning must be used to conduct continuous training, thereby advancing the models further. In order to do so, trustworthy data must be provided through the use of a CAE with a high degree of accuracy.

To re-emphasize , CAE simulation will eventually become a part of artificial intelligence. And as CAE provides data, it plays an important role in advancing artificial intelligence, so it will likely remain a part of artificial intelligence forever after.

Written by Taero Cha (Director of China Business Division)