G13
Reduced Order Modelling

Back to overview

15:30
conference time (CEST, Berlin)
Deep-Learning from Raw, Heterogeneous Engineering Data in the Cloud
27/10/2021 15:30 conference time (CEST, Berlin)
Room: G
P. Baqué (Neural Concept Ltd., CHE); L. Miroslaw, A. Jean (Microsoft, CHE)
P. Baqué (Neural Concept Ltd., CHE); L. Miroslaw, A. Jean (Microsoft, CHE)
Design optimization has become of primary importance for the industry over the last decades. Nevertheless, since a simulation must be re-ran each time an engineer wishes to change the shape being designed, this makes the engineering process slow and costly. A typical approach is therefore to test only a few designs without a fine-grained search in the space of potential variations. In this work, we show how a new generation of AI-inspired methodologies, based on Geometric Convolutional Neural Networks (GCNNs), empowers the engineer with a new approach, to build surrogate models of numerical solvers. We will also demonstrate how the surrogate models are being used in the multiple industry scenarios to accelerate the design cycles. With traditional implementations of machine-learning in engineering, the models were trained on a specific parametric representation of the design space. Therefore, any modification in the design space or any change in the boundary conditions would require generating a new dataset of simulations and training an independent model. On the opposite, GCNN is agnostic to the shape parameters as it processes directly the mesh representation of the design. Hence, optimization or design parameters are decoupled from the learning problem and a single predictor can be trained with a large amount of data and used for many optimization tasks. Practical solutions are already available for teams who want to access such tools . For instance, Neural Concept Shape (NCS), is a software platform that lets engineers, at all levels of expertise, implement the latest deep-learning based engineering practices into their development processes. Using Neural Networks, the simulation engineers and the design engineers can collaborate more efficiently, save costly iterations between teams. Since NCS exploits directly the results of traditional simulations, the integration of HPC clusters with Kubernetes, distributed training on multiple GPUs and the efficient exchange of large volume of data is a force-multiplier for AI. Therefore, NCS has been built to leverage the flexibility and power of cloud HPC services (such as Microsoft Azure) that provide seamless integration with large volumes of simulation results, access to heterogenous computing platforms and ability to deploy trained models in a secure environment for remote engineering and design teams working in different locations. In short, the association of GCNNs with cloud-based HPC is breaking a new barrier in terms of possibilities, performance and convenience for deploying AI-based design optimization at large organizations and provides significant reduction of design times thanks to high performance in simulation, training and inference.
Deep-learning, HPC, Geometric Convolutional Neural Networks, Optimization, Cloud Computing
15:50
conference time (CEST, Berlin)
Interactive Design Space Exploration and How to Make it Happen
27/10/2021 15:50 conference time (CEST, Berlin)
Room: G
M. Bauer, J. Lohse (Navasto GmbH, DEU)
M. Bauer, J. Lohse (Navasto GmbH, DEU)
The digitalization of development processes in high-tech industries has taken up pace, fueled for years by rising High Performance Computing (HPC) capacity and accelerated by the recent pandemic, that drove companies to enable their engineers to work remote with simulations instead of performing physical experiments on site. To compete at the forefront of innovation, these companies are forced to make the most efficient use of their computational power and provide their developers with the tools to exploit the vast amount of data generated in simulation processes. In this contribution, we demonstrate three aspects of a data-driven digitalization strategy on the example of an automotive vehicle design process: We present a tool for interactive design space exploration, quantify the benefits of using adaptive sampling strategy over one-shot DoE approaches and speak on the analysis of parameter importance based on sampled data. Interactive design space exploration exploits Reduced Order Models (ROM), such as Proper Orthogonal Decomposition (POD) and Isomap, as well as surrogate modelling for the real time prediction of CFD solutions, that would otherwise require of the order of days to compute. Together with an intuitive graphical user interface, this enables engineers to quickly define, visualize and evaluate promising design solutions from an infinite number of variants in the parameter space. The use of adaptive sampling (AS) in the data generation process allows to improve model quality within a given computational budget or to reduce the number of required high fidelity simulations without deteriorating model quality. Different adaptive sampling strategies will be presented and evaluated based on ROM and surrogate model error. Finally, we will illustrate how algorithmic determination of the importance of parameters with respect to performance values, such as the drag coefficient, enables developers to make informed choices on where to best invest further effort and how different parameters are coupled.
real-time CFD, Reduced Order Modelling, Surrogates, interactive Design, adaptive sampling, machine learning
16:10
conference time (CEST, Berlin)
Bringing Electrical and Thermal Design Together with Reduced Order Models
27/10/2021 16:10 conference time (CEST, Berlin)
Room: G
B. Blackmore (Siemens Digital Industries Software, CAN)
B. Blackmore (Siemens Digital Industries Software, CAN)
The coupling of power dissipation and temperature is well known in electronics, yet electrothermal effects are rarely considered in the architectural stage of new product design. Incompatibilities between thermal and electrical design tools, and a lack of an electrothermal model supply chain, have proven to be effective barriers to electrothermal analysis in early stage design. A common issue in early thermal design with computational fluid dynamic (CFD) tools is simulating with worst case steady-state power values, and with little or no information about transient power sequencing for even the dominant operational functional goals of the product. The delayed availability of reliable power values often produces over-designed, expensive thermal solutions, or relegates thermal design to a remedial task. On the electrical side, it is common to simulate conceptual circuits with an isothermal assumption. By definition, this means operational temperature impacts on electrical performance are not captured. Temperature related issues with signal timing, current spikes, power levels, control systems, etc. could easily be missed or caught late in the process necessitating expensive re-designs. In order to address these issues, accurate thermal capabilities need to be introduced early in design areas that have traditionally been solely in the realm of electrical engineering. This has not been possible previously as thermal is very much a 3D domain while EDA tools are dominated by 1D toolsets. To bridge this divide, this presentation will illustrate how to leverage the FANTASTIC reduced order model extraction method to convert 3D thermal simulation models into BCI-ROMs in the IEEE standard VHDL-AMS format, and then connect them to form accurate electrothermal circuits in 1D tools. This approach enables power dissipation and other electrical parameters to solve concurrently with temperature, improving the understanding of in-situ product performance earlier in the design process, and ultimately enabling higher quality products to be designed.
Thermal Electrical Co-Design ROM Reduced Order Model
16:30
conference time (CEST, Berlin)
Model-based Integration of Flow and Acoustic Performance Optimization for Axial Fans in Building HVAC Systems
27/10/2021 16:30 conference time (CEST, Berlin)
Room: G
C. Luzzato, B. Afsharpoya (Dassault Systèmes, USA); D. Lauzon (Dassault Systemes, CAN)
C. Luzzato, B. Afsharpoya (Dassault Systèmes, USA); D. Lauzon (Dassault Systemes, CAN)
Heat ventilation and Air Conditioning (HVAC) systems are used in 87% of homes in the US, and represent on average 12% of household expenditures. This breadth of usage means that even small changes in technology can have a huge impact on the footprint of HVAC systems on our environment. Such an impact has led to the progressive ban of Ozone depleting refrigerants in chiller circuits following the Montreal Protocol and Kigali Amendment agreements. This change is affecting the performance of the chiller circuit, and therefore motivating a re-dimensioning or re-design of the full HVAC system. Aside from the chiller circuit, fans themselves are typically responsible for 34% of the average energy consumption of HVAC systems, and are therefore a good subject of work when targeting improved overall system efficiency. Hence, effective design of efficient fans would require iterating towards a better design, typically leveraging a three-dimensional simulation of a rotating geometry. Furthermore, the performance of these fans is highly dependent on the integration to the full HVAC system. In the US, the Spawn-of-EnergyPlus engine is even encouraging HVAC manufacturers to go beyond the design of the HVAC itself, and verify the building integrated HVAC performance directly to minimize energy consumption. Tackling these types of issues often requires representing the HVAC system performance as a set of 1D reduced-order-model nodes. Meeting these types of energy efficiency objectives requires a new approach to system integrated component design. In this paper, we will present aspects of such a design loop, focusing on the aerodynamic and aero-acoustic CAD driven optimization of an axial fan In a second part of the paper, we will include the impact of the fan design in the overall HVAC system performance, represented as a reduced order 1D model. We will show that such development can be organized effectively using a centralized platform, ensuring all key requirements are met.
MBSE, Modelica, CFD, HVAC integration, IEQ, energy efficiency, SEER
×

[TITLE]

[LISTING

[ABSTRACT]

[DATE]

[ROOM]

[KEYWORDS]