Program of our Summer School on Machine Learning for Process and Systems Engineering
Monday, September 27, 2021: Theory of Machine Learning
8:15 - 8:30
Opening of the Summer School
Kai Sundmacher, Max-Planck-Institute, Magdeburg, Germany
8:30 - 9:30 Machine Learning – a (biased) overview in a Nutshell"
Anton Savchenko, Otto-von-Guericke University, Magdeburg, Germany &
Rolf Findeisen, Technische Universität Darmstadt, Germany/ Otto-von-Guericke University Magdeburg, Germany
9:30 - 9:50
Coffee break
9:50 - 11:10
"Basic Principles of Machine Learning: Regression and Classification"
Jens Bremer, Max-Planck-Institute Magdeburg, Germany &
Edgar Sanchez-Medina, Otto-von-Guericke-University Magdeburg, Germany
11:10 - 11:30
Coffee break
11:30 - 12:50 "Introduction to Computational Intelligence Methodologies: Recent Advances in Evolutionary Multi-Objective Optimization"
Sanaz Mostaghim, Otto-von-Guericke University Magdeburg, Germany
12:50 - 14:00
Lunch break
14:00 - 15:30
"Physics-Informed Learning for Low-Dimensional Nonlinear Dynamical Systems Using Operator Inference"
Pawan Goyal & Peter Benner, Max-Planck-Institute Magdeburg, Germany
15:30 - 16:00
Coffee break
16:00 - 17:30
Gaussian processes for dynamical systems, control, and optimization“
Anton Savchenko, Otto-von-Guericke University, Magdeburg, Germany &
Rolf Findeisen, Technische Universität Darmstadt, Germany/ Otto-von-Guericke University Magdeburg, Germany
Tuesday, September 28, 2021: Implementation and Algorithms in Machine Learning
9:00 - 10:30
"Introduction to AI in MATLAB"
Julia Hoerner, Mathworks, Cambridge, United Kingdom &
Philip Laserstein, Mathworks, Ismaning, Germany
10:30 - 11:00
Coffee break
11:00-12:00
"Deep Dive into AI with MaTLAB"
Julia Hoerner, Mathworks, Cambridge, United Kingdom &
Philip Laserstein, Mathworks, Ismaning, Germany
12:00 - 14:00
Lunch break
14:00 - 15:30
"Data-Driven Optimization in Dynamic Systems: Solutions, Structure, and Control "
Julius Martensen & Sebastian Sager, Otto-von-Guericke University, Magdeburg, Germany
16:00 - 22:00
Social event incl. conference dinner
(unfortunately for IMPRS members & speakers only)
Wednesday, September 29, 2021: Machine Learning Applications in Process and Systems Engineering
9:00 - 10:30
"From neural networks to ordinary and partial differential, and back"
Thomas Richter & Christian Lessig, Otto-von-Guericke University Magdeburg, Germany
10:30 - 11:00
Coffee break
11:00 - 12:30
„Machine learning approaches for the prediction of oxygen evolution catalysts“
Stefan Palkovits, Rheinisch Westfälische Technische Hochschule Aachen, Germany
12:30 - 14:00
Lunch break
14:00 - 15:30
"Machine Learning and Hybrid Modeling in Process Engineering"
Moritz von Stosch, datahow, Dübendorf, Switzerland
15:30
Closing remarks
..............................................................................................
Abstracts of the lectures
“Machine Learning – a (biased) overview in a Nutshell"
Anton Savchenko, Otto-von-Guericke University, Magdeburg, Germany &
Rolf Findeisen, Technische Universität Darmstadt, Germany/ Otto-von-Guericke University Magdeburg, Germany
This lecture provides an overview and introduction of machine learning approaches with a focus towards dynamical and chemical engineering systems. We outline, introduce and classify different approaches and their use, spanning from regression and classification to support vector machines, Gaussian processes, and Neural Networks.
"Basic Principles of Machine Learning: Regression and Classification"
Jens Bremer , Max-Planck-Institute Magdeburg, Germany &
Edgar Sanchez Medina, Otto-von-Guericke University Magdeburg, Germany
The lecture deals with the emerging topic of machine learning in chemical engineering. You are familiarized with typical problem types such as logistic regression, classification and support vector machines and enabled to apply different models/estimators. With these methods in mind, different examples from chemical engineering are used as an illustration of the methods. In summary, you are ready to analyze large research data in a new way and assess the use of either a mechanistic or data-driven model or a combination of both.
"Introduction to Computational Intelligence Methodologies: Recent Advances in Evolutionary Multi-Objective Optimization"
Sanaz Mostaghim, Otto-von-Guericke University Magdeburg, Germany
This talk will present an overview about the field of Computational Intelligence which is the origin of AI and Machine Learning (ML) methodologies. After an introduction into the field, the main focus is on the class of multi-objective optimization algorithms which are getting popular in various disciplines. The novelty of such algorithms in contrast to the existing ML algorithms is that they can learn several functions at the same time. The talk will provide an overview about various classes of multi-objective problems with relevance to the chemical and process engineering field.
"Physics-Informed Learning for Low-Dimensional Nonlinear Dynamical Systems Using Operator Inference"
Pawan Goyal & Peter Benner, Max-Planck-Institute Magdeburg, Germany
Dynamical modeling of a process is essential to study its transient behavior and to perform engineering tasks such as control and optimization. With the increased accessibility of data, learning models directly from the data is receiving significant attention. In the first part of the presentation, we will describe the historical development up to the current activities. Here, we will focus mostly on linear systems. In practice, it is often desirable to construct compact low-dimensional models describing complex nonlinear dynamics, allowing simulation, optimization and control of engineering systems on modest computer hardware. The second part of the lecture introduces the data-driven operator inference (OpInf) approach for learning low-dimensional models of nonlinear systems. Here, it is assumed that the structure of the non-linearity is known only at an abstract level. In light of this, we discuss how to tailor the OpInf approach to get interpretable physical models. Moreover, we discuss the usage of deep learning to improve the performance of the approach further. We illustrate the methodologies to learn low-dimensional dynamical models by means of often encountered engineering problems.
"Gaussian processes for dynamical systems, control, and optimization“
Anton Savchenko, Otto-von-Guericke University, Magdeburg, Germany &
Rolf Findeisen, Technische Universität Darmstadt, Germany/ Otto-von-Guericke University Magdeburg, Germany
This lecture provides a brief introduction into the fundamentals of Gaussian process for dynamical systems and control. After a recap of support vector machines, we focus on the fundamentals of Gaussian processes and how they can be used towards reference and model learning, as well as process optimization and control. The derivations are underlined by various chemical engineering and control examples.
" Introduction to AI in MATLAB"
Julia Hoerner, Mathworks, Cambridge, United Kingdom &
Philip Laserstein, Mathworks, Ismaning, Germany
- Definition of AI
- AI workflow
- Fundamentals of Machine Learning and Deep Learning (incl 2 examples)
- Summary
"Deep into AI with MATLAB "
Julia Hoerner, Mathworks, Cambridge, United Kingdom &
Philip Laserstein, Mathworks, Ismaning, Germany
- Quick intro
- Hyperparameter tuning
- Customisation of networks
- Interoperability with other frameworks
- Summary
"Data-Driven Optimization in Dynamic Systems: Solutions, Structure, and Control"
Julius Martensen & Sebastian Sager, Otto-von-Guericke University, Magdeburg, Germany
Machine Learning (ML) has proven to be a valuable tool for data analysis, control, and many other areas. Modeling and optimization are at the core of extracting patterns, finding anomalies, and generating additional scientific insight from experimental data. In this lecture, we will present applications of ML in connection to the fields of mixed-integer optimization, neural networks, and system estimation. The use of optimal control techniques within the modelling and training of ML will be explored and explained by showcasing the classification of cardiac arrhythmia. Furthermore, we will highlight some of the latest work in the field of symbolic regression and how to apply it to generate the equations of motion from data.
"From neural networks to ordinary and partial differential, and back"
Thomas Richter & Christian Lessig , Otto-von-Guericke University Magdeburg, Germany
We discuss connections between (ordinary and) partial differential equations and neural networks. One the one hand, the theory of these equations can be used to gain theoretical and practical insight into neural networks. On the other hand, neural networks also provide novel tools for solving partial differential equations. In this lecture, we will discuss both sides of this connection and how mathematics can be used in this context.
"Machine learning approaches for the prediction of oxygen evolution catalysts"
Stefan Palkovits, RWTH Aachen, Germany
The oxygen evolution reaction (OER) is currently under intense investigation because of its relevance for a circular economy. The lecture will show how to gather knowledge about the the OER based on literature data. A Python workflow will be used showing the steps from data pretreatment over hyper parameter search up to the final Machine Learning predictions. These predictions will be also correlated to actual laboratory data.
"Machine Learning and Hybrid Modeling in Process Engineering"
Moritz von Stosch, datahow, Dübendorf, Switzerland
The advent of new machine learning algorithms and more powerful computers has rendered the solution of problems possible that were for long considered not to be machine solvable. Also in the process industries these tools have found increasing attention and application. However, machine learning tools require high quantities of qualitative data, the generation of which being typically expensive. In contrast to other industries, fundamental knowledge in form of mathematical equations is available in the process industries. By combining the fundamental knowledge with machine learning tools in so called hybrid models, the data requirements can be significantly reduced and the extrapolation properties of the hybrid models remain valid beyond experimentally tested combinations. This lecture provides an introduction into machine learning and hybrid modeling and their application in the industries.