The Greydient ESRs

The Greydient project is dedicated to developing solutions to the pressing challenges we face in integrating data and models into self-learning systems that enable the intelligent design, optimization and control of products, processes and systems. To tackle these challenges, Greydient will train 15 Early Stage Researchers (ESRs) in a wide spectrum of fields, including the modelling, propagation and quantification of the relevant variabilities, the application of big data and machine learning methods, as well as the optimal combination of data-driven approaches with numerical models. The ESRs are at the center of this project. Each will obtain a PhD from an internationally respected university, build experience in communicating and disseminating their work, applying their research skills in a non-academic context and receive in-depth training in transferable skills such as commercialization, collaboration and entrepreneurship. The recruitment was incredibly successful and we are very proud to present the 15 new faces of the project!

Thermal modelling and online heat transfer estimation using Active Learning grey-box approaches

Institution: KU Leuven, Belgium

Objectives:

  1. Construct a production-process virtual twin consisting of an ensemble of data-driven black box approaches, surrogate models and full-scale, non-linear, finite-element computations
  2. Apply active learning to train this virtual twin based on the full-scale white box model
  3. Apply this virtual twin in combination with active learning to create an optimal sensor lay-out
  4. Apply the developed methodology on a projection welding machine at KU Leuven.

Expected results: A generic methodology to optimize large-scale sensor arrays in production plants, as well as perform online process control, via an optimal combination of simulation models with experimental data, validated on the projection welding process.

Bio: Hey! I’m Miriam, doing my PhD with the KU Leuven about performing reliable process monitoring considering uncertainties. Really excited about my research and collaborating with the great people from this project. Outside of work, I enjoy adventuring just as much as a cozy day. m

 
Digital Image Correlation enriched quantification of polymorphic random field uncertainty in finite-element models Institution: KU Leuven, Belgium Objectives:
  1. Develop a methodology to simulate strain fields based on imprecise random-field valued material parameters using a Digital Image Correlation simulator
  2. Apply non-linear dimension-reduction approaches to deduce a low-dimensional manifold from experimental strain fields
  3. Train a physics-informed deep-learning architecture to predict imprecise random-field descriptors including expert knowledge
  4. Validation of the inverse methodology on functionally graded materials.
Expected results: A methodology to inversely quantify non-homogeneous random-field uncertainty based on full-field strain measurements. Illustration of the method on the identification of functionally graded materials (FGMs). Bio: I’m a 24 years old french engineer specialised in computer science, tennis player and mountain lover. Passionate about image processing, I will be working during my PhD on the Digital Image Correlation technology which allows to characterize locally and precisely a material.

Efficient fuzzy analysis for real-time noise and vibration analysis

Institution: Leibniz Universität Hannover, Germany

Objectives:

  1. Expand the concept of intervening variables towards explicit dynamic FE analysis
  2. Combine this approach with series-expansion methods to solve the fuzzy-propagation problem
  3. Apply these methods to a practical car-crash analysis.

Expected results: Algorithms for real-time characterization of imprecise dynamical response, applied to transient dynamic analysis in the context of noise and vibration analysis.

 

Fast imprecise surrogate model for medium-fidelity data analysis

Institution: Leibniz Universität Hannover, Germany

Objectives:

  1. Develop a highly efficient and effective surrogate model to derive predictions on the structural behaviour from a mix of medium-fidelity data
  2. Benchmark the Interval Neural Network with respect to other set-valued surrogate models such as interval predictor models or probabilistic Neural Networks.
  3. Apply Interval Neural Networks in an OpenCOSSAN implementation on a practical car-crash analysis in collaboration with SCALE.

Expected results: Algorithms for fast data analysis and decision support.

Bio: Hi, I’m Chen Ding from China. I am a PhD student at the Leibniz University Hannover, working on “Fast imprecise surrogate model for mid-fidelity data analysis”. In my spare time, I love to record my life, learn Japanese from Animation, and immerse myself in music.

Advanced probabilistic black-box modelling methods for reliability, risk and resilience analysis of smart critical systems

Institution: Politecnico di Milano, Italy

Objectives: To develop a simulation framework for smart critical system resilience analysis that:

  1. Identifies component contribution to the overall system risk
  2. Prioritizes system improvement decisions for accident recovery
  3. Can account for inter-dependence in the system’s components 
  4. Can account for a wide range of both small and large uncertainties.

Expected results: A set of tools comprising advanced AI black-box models and advanced Monte Carlo simulation for the reliability, risk and resilience analysis of smart critical systems and infrastructures.

Bio: I am Leonardo, a polimi PhD student and a curious person by nature. I declare myself as a true lover of math, literature, films, music, and related arts of any kind. We are currently working full-time on the Greydient Project for a better, safer, and more reliable world 😀

 

Multi-fidelity grey-box models for integrated, condition-informed, dynamic risk assessment

Institution: Politecnico di Milano, Italy

Objectives:

  1. Develop dynamic modelling approaches capable of modelling both hard (physical) and logic (cyber) systems failures
  2. Construct a multi-fidelity grey-box model for dynamic reliability assessment of systems 
  3. Adopt a Value of Information (VoI) approach to identify the most appropriate multi-fidelity grey-box model
  4. Validate the developments for a realistic case study in collaboration with Aramis.

Expected results: A validated dynamic multi-fidelity grey-box approach for risk assessment based on a VoI approach, including physical and cyber-systems failures.

Bio: I’m a challenge-driven electrical engineer looking for new mathematical and computational tools to make our world safer, greener, and happier. I love learning from diverse groups and living life to the fullest. You can also call me a maker, musician, or vegan home cook!

 

Optimization of Energy Networks Equipped with Prognostics and Health Management capabilities

Institution: Aramis, Italy

Objectives:

  1. Development of a methodology to estimate the reliability and availability of energy-grid systems including components equipped with Prognostics and Health Management (PHM) capabilities
  2. Estimate the reliability and availability of energy grid systems including components equipped with Prognostics and Health Management (PHM) capabilities (based on (1)) 
  3. Apply the developed methods to obtain optimal solutions depending on the O&M costs as well.

Expected results: A set of methodologies and algorithms that have to potential to significantly improve the existing approaches for energy-grid optimization.

A Learning-based Flexibility Approach for Lack-of-Knowledge Robustness in Early-Phase Crashworthiness Design

Institution: Technische Universität München, Germany

Objectives:

  1. Derive artificial intelligence (AI) approaches for adaptively maximizing the flexibility of car-crashworthiness design based on low-fidelity component-level models
  2. Use AI to link experimental data and high-fidelity FE models to enhance the representation of mechanical characteristics provided by low-fidelity models
  3. Apply these methods at a practical case in collaboration with SCALE GmbH.

Expected results: Extension of the Solution Space Method (SSM) to derive optimized crash structures to account for lack-of-knowledge and adapt successively via machine-learning methods. Given the power of machine-learning capabilities, more complex crash load cases can be addressed in the future to enable an efficient and tailored virtual product development where experimental or computational data can be integrated.

Bio: Whether it is in the mountains or in the office, I like pushing boundaries to reach new horizons. If this brought me on top of beautiful mountain ranges, it will now allow me to tackle the difficult path of improving an early-stage design method for robust crashworthiness at TUM.

Grey-box Approach for Verification & Validation (V&V) for Crashworthiness – Learning-based validation of crash models

Institution: BMW, Germany

Objectives:

  1. Derive artificial intelligence (AI) methods to allow validation for crash cases when very limited or even no repetition of physical tests is possible
  2. Derive grey-box approaches to support the engineer improving and interpreting numerical simulations for crashworthiness by identifying correlations, weaknesses, and shortcomings of the virtual models
  3. Application of these tasks at a practical case at BMW including Digital Image Correlation.

Expected results: As a result of the proposed developments, a new approach for Verification & Validation will be established where no stochastic physical experiments are required, and better and more confirmed evidence of the virtual models is achieved.

Bio: I am Giada Colella, a passionate and strong-willed mechanical engineer. In my daily life, I like to be inspired by the people that surround me. I love stepping out of my comfort zone working in heterogeneous teams. My motto: “A smooth sea never made a skilled sailor”.

Learning multivariate time series – Application to tire manufacturing and wear prediction

Institution: Sigma-Clermont, France

Objectives:

  1. Develop an approach to learn time-univariate time series,
  2. Extend this method to the learning of cross-dependent multivariate time series
  3. Apply this method to industrial problems in collaboration with Michelin.

Expected results: General method able to learn multivariate time series from data with deep neural network or other suitable machine learning techniques. Focus on time-series forecasting. Potential application to the inference of unobserved time series. Representative examples in tire industry.

Bio: My name is Thabang Lebese, born and raised in Johannesburg, South Africa. I received my training in pure and applied mathematics. I have an interest in artificial intelligence (AI), I previously worked on different applications of machine learning (ML) including NLP, neural-symbolic AI, applications in high energy physics and modelling of climate and conflict. My current PhD research focuses on grey-box modelling using multivariate time-series for tire manufacturing and wear prediction. I enjoy global politics, listening to music and taking a “Sho’t Left” every now and then for replenishment. I consider Nelson Mandela to be the most outstanding leader of our time and my favorite comedian is Dave Chapelle.

The use of machine learning and grey-box models to solve complex time-consuming RBDO problems

Institution: Sigma-Clermont, France

Objectives:

  1. Develop an efficient approach for reliability estimation based on machine-learning separators
  2. Apply this method in a multi-fidelity modelling context
  3. Apply zero-order optimization algorithms to solve a reliability-based design-optimization problem based on machine-learning separators
  4. Apply these developments to a realistic case study of a production process in collaboration with Radiall, a producer of electric connectors for automotive industry.

Expected results: Efficient and new methodology for RBDO problems based on machine learning separators, multi-fidelity modelling and zero-order optimization algorithms to efficiently perform a reliability-based design-optimization of production processes.

Bio: 27 years old, Italian, MSc in Structural Engineering, charmed by UQ world. My research focuses on Machine Learning separators and Surrogate Modeling to enhance Reliability-Based Design. On the dark side of the Moon I’m a drummer and a tennis player. Addicted to Russian literature

Random-field modelling of the mechanical properties of composite materials at different scales

Institution: Aristotle University of Thessaloniki, Greece

Objectives:

  1. Reconstruct the morphology of random microstructures based on computer-simulated or SEM images
  2. parameter identification of random fields describing the spatial variability of mechanical properties of composites at different scales
  3. Investigation of the effect of the probabilistic characteristics of the input random fields on the response variability of composite structures using stochastic finite elements
  4. Application of the methodology to design reliable composite materials with the desired macroscopic properties based on a tuned microstructure.

Expected results: A generic approach to identify and reconstruct random fields on mechanical properties, as well as to assess the effect of spatial variability on structural reliability of composite components, based on a numerical model of the structure, advanced probabilistic techniques and microstructure reconstruction from SEM images.

Bio: I am Panagiotis Gavallas, a Civil Engineer with a Master’s in Computational Mechanics from TUM. My research will be focused on the mechanical behaviour of composite materials at multiple scales using random fields. I like being close to nature and am open to new experiences.

Active-learning multi-fidelity grey-box modelling

Institution: ETH Zurich, Switzerland

Objectives:

  1. Construct multi-fidelity models based on models with different levels of discretization, linearization, etc.
  2. Use active learning to optimally distribute the computational budget among different fidelity levels
  3. Include data-driven approaches in the multi-fidelity framework and use active learning to select from the combination of white-box, data-driven models
  4. Validate the developed methodologies on a case study at EDF.

Expected results: An algorithm to optimally distribute the available computational/financial resources between different levels of modelling fidelity and different surrogate modelling strategies will be implemented. The resulting strategy is expected to provide a significant improvement on the feasibility of complex uncertainty quantification analyses for expensive grey-box-type models.

Bio: Electrical and computer engineer from Greece. Having worked for 4 years in the UK as an ML engineer, I recently moved to Zurich to start my PhD. Really excited to work on multifidelity grey-box modelling as part of GREYDIENT!
Things I love: Maths, beach volleyball, travelling.

 

Uncertainty quantification and reliability analysis for noisy grey-box models

Institution: ETH Zurich, Switzerland

Objectives:

  1. Re-formulate the reliability problem in the context of noisy grey-box models
  2. Apply sparse polynomial chaos expansions as a denoising tool 
  3. Combine denoising and active learning to obtain highly efficient reliability estimates for grey-box models based on expensive industrial models and noisy data.

Expected results: A method that provides reliable and cost-efficient estimates of the probability of failure of structures and systems that, due to employed grey-box modelling approaches, contain noise-contaminated limit-state functions (e.g., due to numerical or experimental noise).

Bio: Hey! I am Anderson, a Brazilian civil engineer currently based in Switzerland. My research is about developing ways to cope with noise and aleatory variability present in grey-box models. Off-the-clock, I like cooking while listening to good music and chatting with friends.

Machine learning and physical modelling interactions for vehicle batteries optimisation

Institution: Phimeca, France

Objectives

  1. Develop a multi-physics numerical model to represent battery degradation and performance
  2. Implement this model in combination with real monitoring data in a functional mock-up interface (FMI) framework
  3. Develop a method to iteratively update the numerical model via Bayesian model updating and 
  4. Validate this method on a real electric vehicle battery pack case.

Expected results: A method for efficiently and effectively monitoring battery degradation and performance using an adaptive grey-box virtual twin approach, based on a FMI integration of a multi-physical model with online monitoring data.

Bio: Ciao! I’m Francesco, coming from Milano to work at Phimeca Engineering in Paris! I joined Greydient because is the perfect chance for performing cutting edge research with an amazing group of international people. A ball, a beer and a bunch of friend: here lies my happiness!

Greydient Consortium

Greydient brings together experts from academy and leading European companies: 6 countries, 7 universities and 11 industrial partners.