by Hamid Asgari, Juha Kortelainen, and Mikko Tahkola (VTT)

Artificial intelligence, machine learning and artificial neural networks are introducing interesting opportunities to engineering design as well as to monitoring and operations of systems and processes. Core and enabling technologies are evolving fast, with great potential for industrial applications. Machine learning, combined with simulation, can enable models that are both fast and sufficiently accurate to be used in new applications.

Modelling and simulation are commonly used tools in mainstream engineering design. For instance, numerous open source and propriety software applications are available for structural analysis, computational fluid dynamics, and system simulation. These engineering tools have revolutionised the design process, enabling increasingly demanding designs to be completed more quickly and efficiently. In engineering design use, the performance of modelling and simulation tools is becoming increasingly important. The use of large design studies and application of mathematical optimisation are emphasising the computational performance of simulations, especially when thousands of simulations are needed. Increased performance of each simulation can remarkably improve the overall analysis time and enable new information to be gained faster. This means that larger, more thorough studies can be completed within a given time.

The current interest in digital twins has also emphasised the need for faster simulations and improved computational efficiency. A digital twin is a digital, i.e., computer-based, replica of a real-world product, system or process. The digital twin represents the relevant features (from a use and operation perspective) of the real-world twin and synchronises its state with it. One way to define and categorise digital twins is to contemplate their ability to represent the state of the real-world system in time. If the digital twin is able to describe the history and present state of the real twin, but cannot predict its future state, the digital twin does not necessarily need any simulation features and can be considered a “descriptive digital twin”. On the other hand, if the digital twin can predict the future state of the real twin, some simulation capabilities are usually needed and the digital twin can be categorised as a “predictive digital twin”. Furthermore, if the digital twin is used to optimise the function and operation of the real twin, some mathematical optimisation capabilities are generally needed, and we can categorise it as a “prescriptive digital twin”. The more prescriptive the use of a digital twin, the more its computational efficiency matters. Thus, innovative ways to simulate complex systems, especially involving complex physical phenomena, are needed.

A surrogate model is a simplified and usually computationally efficient replacement of the original or more accurate model of the target. The original model can be based on physics-based simulation, which can require considerable computing resources and be very time consuming. A solution may lie in machine learning and artificial neural networks, which are currently being researched and developed to produce data-driven models.

Machine learning is a subset of artificial intelligence that deals with exploring the data structure and fitting the data into a model. It relies on the use of computers, algorithms, and data processing techniques for clustering, regression, classification, and pattern recognition. Advances in artificial intelligence and machine learning have helped improve engineering techniques in different disciplines in ways that may not have been possible with conventional methods. Machine learning algorithms are grouped into three main categories: supervised learning, unsupervised learning, and reinforcement learning. Methods utilising so-called deep artificial neural networks are commonly called deep learning. Deep learning is employed to set up a complex artificial neural network structure with multiple layers to train big datasets with complex interconnection relationships. Figure 1 illustrates a classification of machine learning algorithms and some of their applications [L1]. ]. Figure 2 represents applications of machine learning in industry for fault detection, prediction and prevention [1]. As Figures 1 and 2 show, machine learning may be employed in a variety of engineering applications, such as process optimisation, predictions, diagnostics, big data visualisation, and robot navigation.

Figure 1: A classification of machine learning algorithms [1].

Figure 1: A classification of machine learning algorithms [1].

Figure 2. Applications of ML in industry for fault detection, prediction and prevention [2].
Figure 2: Applications of ML in industry for fault detection, prediction and prevention [2].

Artificial neural networks have been successfully used in various applications across different domains and can be applied to surrogate modelling as well. A range of open source software for building artificial neural network models is available. The surrogate approach is model agnostic in the sense that only the data from the simulation software is required. Depending on the simulation speed, it can be useful to be able to parallelise the data generation to gather the data faster. Similarly, it is possible to speed up surrogate model development by parallelising the artificial neural network training process.

We investigated the use of artificial neural networks to create a surrogate model of a physics-based industrial process simulation model. Modelled systems were a liquid level-controlled water tank process, and a methanation reactor in a power-to-gas process, shown in Figure 3 [2]. In the former case, normalised root mean squared error of the surrogate model was 0.01% and in the latter about 4%. In the electrical machine domain, we have created surrogate models of a permanent magnet synchronous motor that are about 2,500 times faster than the original model, with a normalised root mean squared error of between 0.5 and 10%, depending on the operating point. The case studies show that surrogate modelling is a potential tool to enhance research, development and design work efficiency by enabling faster simulation.


Figure 3: Physics-based simulation model of the methanation reactor in a power-to-gas process [3].
Figure 3: Physics-based simulation model of the methanation reactor in a power-to-gas process [3].



[1] A. Angelopoulos, et al.: “Tackling Faults in the Industry 4.0 Era—A Survey of Machine-Learning Solutions and Key Aspects”, Sensors, Basel, 20(1): 109, 2020.
[2] M. Tahkola: “Developing dynamic machine learning surrogate models of physics-based industrial process simulation models”, master’s thesis, University of Oulu, 2020.

Please contact:
Hamid Asgari, Juha Kortelainen, Mikko Tahkola, VTT, Finland
This email address is being protected from spambots. You need JavaScript enabled to view it., This email address is being protected from spambots. You need JavaScript enabled to view it., This email address is being protected from spambots. You need JavaScript enabled to view it.

Next issue: July 2023
Special theme:
"Eplainable AI"
Call for the next issue
Image ERCIM News 122 epub
This issue in ePub format

Get the latest issue to your desktop
RSS Feed
Cookies user preferences
We use cookies to ensure you to get the best experience on our website. If you decline the use of cookies, this website may not function as expected.
Accept all
Decline all
Read more
Tools used to analyze the data to measure the effectiveness of a website and to understand how it works.
Google Analytics
Set of techniques which have for object the commercial strategy and in particular the market study.
DoubleClick/Google Marketing