Introduction of Reduced Order Modelling and its Application for Model based real-time Optimization
The cost of computing and simulation has become an important component of the total design cycle. In particular, anything related to iterative computing such as optimization, reliability or robustness studies requires frequent relaunch of models. This is costly even for one design or simulation, due to the fact that any sensitivity related analysis requires detailed models, sufficiently sensitive to changes of model variables (endogenous) or parameters (exogenous). The associated results are simulation outputs, experimental measurements or both while the primary purpose of simulation models is to replace real-life experiments by lower cost and easier to control counterparts. Unfortunately, the desired simplicity of models (often formulated in terms of oversimplified surrogate models) has its limits due to often complex physics or lack of representative models. The introduction of Reduced Order Modelling (ROM), is aimed at solving two problems at a time, namely the requirement for precision accompanied by the constraints on computational effort. Until recently there were no sufficiently accurate and realistic intermediate solutions allowing to learn from experiments or simulations and establish a link between the two, allowing for a “digital twin” approach or simply a “coupled simulation-experimental” model. Additionally, modeling is often confronted with issues related to licensing, multi-scale considerations and confidentiality, especially in an industrial environment. The ROM technique is an innovative and promising solution to advance engineering to yet another limit in terms of what we can call real-time modelling. In recent years the concept of ROM, initiated originally from system modelling, has been introduced and has picked up acceptability and efficiency both in applications and algorithmic maturity. The fundamental idea is to exploit compression techniques borrowed from classical matrix algebra and signal processing, combined with new concepts related to machine learning and image processing, to reduce the computational effort necessary to calculate and exploit existing models and/or experimental results. Given an existing dataset (know-how) on the behavior of a system subject to variations of its internal components or external loadings it is intended to exploit this know-how in form of a black-box in order to predict its behavior without extensive computing, thus enabling re-design, etc.
Who should attend?
Anyone interested in accelerating sensitivity analysis, optimization, reliability, robustness and globally anyone who needs to understand and initiate the application of ML techniques in CAE is a candidate for participating to this seminar. You don’t need any deep understanding of machine learning nor complex mathematics or matrix algebra. Your engineering education is enough, even though experience in simulation techniques (various solvers) and optimization processes (not algorithms) is very useful. The participants can be academics, above MSC level students, designers, architects, engineers or project managers and team leaders at various design or decision-making stages of CAD/CAE based manufacturing. Topics from automotive, aeronautics, medical and civil engineering will be exploited for demonstrations.
- Model Reduction
- Theoretical background
- Eigenvalue problem
- Singular value decomposition
- Methods of Reduction
- Proper Orthogonal Decomposition
- Central Voronoi Tessellation
- Fast Fourier Transforms
- Neural Networks
- Support Vector Machines
- Dynamic Mode decomposition
- Application of Reduced Order Model as a surrogate model
- Reliability & Robustness
- Entropy and Complexity
- Sled Test
Dr. Kambiz Kayvantash
Kambiz Kayvantash has a PhD from University GHS of Essen, Germany, and an executive MBA from HEC, Paris, France. In the past he has occupied various technical and managerial positions both in academia and industry (University of Essen, MECALOG, ALTAIR, Cranfield University/Cranfield Impact Centre, CIVITEC, CADLM/ École spéciale des travaux publics. He is currently the CTO of CADLM and develops Artificial Intelligence based solutions for predictive, real-time industrial applications such as Autonomous Vehicle, Crash and Safety, Health Monitoring, etc.
|06 Feb - 09 Feb 2023||English||1340 EUR (1650 EUR from 10 Jan 2023 )||4066|
|Online (Online-Seminar)||» Register|
Dr. Kambiz Kayvantash (CADLM)
All prices are exclusive of VAT.