Mathematical methods for the optimization of district heating systems and geothermal storages

Abstract

The energy transition and decarbonization of the economy can only succeed if in addition to the electricity production also the heat supply is transformed. The contribution of  renewable energies such as solar heat and geothermal energy has to be increased. This requires a suitable temporal and spatial balance of supply and demand  caused by seasonal and weather-dependent fluctuations. Here, a new generation of heating networks as well as geothermal storages play an important role. 
We investigate modeling, simulation, optimization, and control of heating systems equipped with geothermal energy storages. This helps to understand the complex charging and discharging behavior and to find a suitable shape, insulation and size of these storages, an efficient spatial arrangement of heat exchangers and cost-optimal controls of heating systems. Mathematical problems from shape optimization and stochastic optimal control are solved using methods from numerical analysis and machine learning.

• Mathematical modeling and the development of digital twins for the simulation of the complex charging and discharging behavior of a geothermal energy storages
• Application of mathematical shape optimization to find a suitable shape, insulation and size of these storage systems
• Application of stochastic optimal control methods for the  cost-optimal and anticipatory control under uncertainty  of heating systems equipped with geothermal geothermal energy storages
• Development of numerical and machine learning methods for the solution of the aforementioned problems

https://www.b-tu.de/fg-wirtschaftsmathematik/forschung/projekte/mones

11/2022-10/2025, 3 Jahre

Digital life cycle of hybrid-electric drive systems

Abstract

The aviation industry is an important economic factor and is experiencing continuous growth [1] in passenger and freight traffic, although it faces challenges such as strict emission targets [2] and safety requirements. [3] Emission targets for 2030 are to be reduced by 40% compared to 1990, and it is proposed to increase this target to 55%. Electrification and digitalization are key factors in achieving these targets. [4] The aviation industry needs a realignment and faces a growing need for transparency in terms of components and processes. The introduction of (hybrid) electric drives in aviation is seen as a solution - similar to the automotive industry.
The DIREKT project aims to create and network digital twins for the development, production and maintenance of (hybrid) electric drives in the aviation industry in order to meet emission and safety targets and create economic opportunities in transformation regions such as the Lausitz region.

• Assembly analysis of an additively manufactured replica of an aircraft engine stator component
• First iteration of the prototype AR assistance system and its visualization as a mobile AR application

11/2022-12/2025, 36 months (project start is delayed)

AI-controlled robotic industrial machines

Abstract

The main motivation of the project is bringing AI‐methods into Lusatian SME and particularly into special‐purpose machine construction. This includes the development of solutions—as universal as possible—for controlling industrial robots through a combination of numerical and symbolical AI‐methods. The scenario is settled around small and medium‐sized production series where automation by numerical AI alone is difficult because of the small number of examples and also automation by symbolical AI alone is difficult because of the variety of tasks. The project tries to identify suitable relationships and associations between the two AI flavours and combine them with other methods like simulated training data to resolve those shortcomings.

• cognitive machine controller realising perception‐action‐cycles
• perception‐action‐cycles integrating numerical and symbolical AI‐methods
• symbolical reinforcement‐learning algorithm
• small‐data‐approaches for dealing with little real sensory data

10/2022–09/2025, 36 months

Data-driven process modelling of closed-die forging processes to increase productivity using adaptive tool design methodology

Abstract

Studies have shown that many disturbances in the forging process can affect the quality of forgings. For instance, press and tool deflections can lead to relative displacement of workpieces, which has significant influence on the geometric accuracy of products.
The goal of the project is to use artificial intelligence to design new forging tool work surfaces to improve the resilience of the forging process. For this purpose, features of the process will be generated and patterns defined in the first periode of the project. The type of forging defects as well as the corresponding causes are to be identified using numerical analysis. Besides, a novel sensor network fusing RADAR sensors of 120 GHz is being developed for constructing a data-driven model of the forging process. 

https://gepris.dfg.de/gepris/projekt/520194997

08/2023-08/2026, 36 months

Data- and Response-Surface-Driven Design Assistant for Controlled Flexible Multibody Systems

Abstract

Flexible multibody systems appear in many applications, ranging from industrial robots in manufacturing to robotized support in surgery, usually including actuation and corresponding controllers. Today, such systems are usually designed based on an analysis-centric approach, where human engineers modify the mechanical and control design separately by hand based on manual inspection of simulation and experiment results to find better designs. However, simultaneous design of mechanics and control may better exploit the full optimization potential. The artificially intelligent design assistant to be developed in this joint project with Prof. Peter Eberhard from the University of Stuttgart shall bridge this gap, revolutionizing the design and analysis process of controlled flexible multibody systems by leveraging modern machine-learning techniques at several points in the design process. 

• surrogate-based multicriterion design approach with formal criteria
• AI-strategies for design space reduction
• joint optimization approach for kinematics, kinetics, body shapes and controllers
• partly substitution of conventional causal modeling by data-driven modeling

https://www.itm.uni-stuttgart.de/en/spp_2353/

8/2022-12/2024, 36 months

Digital thread-based engine design with embedded artificial intelligence

Abstract

Expanding the competences for digitally designing turbo engines is a key capability for Rolls-Royce Deutschland (RRD) to maintain its leading position as engine manufacturer in the global market. A deep understanding of the continually growing complexity of modern aero-engines is necessary to improve the propulsion efficiency while reducing development cycle times. The joint project DARWIN provides an ideal platform for a deep cooperation between RRD with the Brandenburg University of Technology Cottbus-Senftenberg (BTU), the Technical University of Darmstadt and the University of Surrey as well as the research partners German Aerospace Center (DLR) and Technical University of Dresden. Especially in the work-package 3.3, the BTU-Chair of Engineering Mechanics and Vehicle Dynamics will develop methods for AI-based, multidisciplinary optimization of aero-engines.
 

• Improvement of efficiency, accuracy and robustness of digital design-applications predicting the multi-disciplinary behavior of engine systems with hybrid and scale-resolving methods.
• Further automation of development processes by integrating digital twins of engine components, including manufacturing variation and wear and tear, as well as accounting for data from development, validation and operating phases of the product life cycle according to the "digital thread" concept.
• Application and enhancement of artificial intelligence and machine learning tools to accelerate design processes and improve the quality and accuracy of analysis results.
• Validation of performance and reliability, thus industrial readiness, of developed design methods via challenging industrial use cases.
   

3 Jahre

Behavior control of intelligent machines

Abstract

The essential added value of intelligent machines lies in their ability to learn independently from experience, to plan, to act sensibly (via actuators) and - in a simple sense - to "think". These aspects are summarized as "behavioral control" of the machine. Since today's technical solutions still show considerable deficits in this respect, we aim to show with the support of the European Regional Development Fund (ERDF) under the name "Behavioral Control for Intelligent Machines" how a cognitive system transforms sensory information (interaction circuit) into semantic information so that it can be processed logically within data Structures that allow higher cognitive abilities, such as coping, and thus a meaning-driven control of the inner cognitive circle (physical system components).We develop a cognitive research and experimentation system that deals with a mouse-maze problem inspired by Claude Shannon’s Theseus, with the aim of explaining the process flows based on a system theoretical and integrative Approach by considering traditional cognitive architectures. For this purpose, we expand the Perception-Action-Cycle (PAC) of cognitive systems to include an inner stage with database semantics and a superstructure including psychologically and biologically motivated functionalities [RGH⁺19]. Golem executes instructions (motor function) and perceives structures and elements of the world through the sensor system; Homunculus is responsible for thinking and imagination about the world (inner stage); Theseus - for strategy finding; Demiurge - for the needs and Argus (as a control authority) - for damage avoidance [MBE⁺21].

• The Cognitive Agent is realized as a mouse in a maze whose behavioral control takes place in a deterministic as well as in a non-deterministic environment simulation.
• The tasks  such as finding objects and exploration are implemented on the common basis of Markov Decision Processes (MDP) efficiently.
• The structured data on the semantic databank and the interaction with the environment lead to an inner stage being created during the exploration phase, on which problem-solving programs are carried out risk-free and thus contributes to damage avoidance and goal assessment.

04/2020–06/2022

Perspectives in Stochastic Optimization and Applications 

Abstract

The project aims to foster a cooperation between mathematicians of BTU and the African Institute of Mathematical Sciences (AIMS) Ghana. The activities comprise research workshops, summer schools and research stays at BTU for AIMS graduates aiming a PhD in mathematics. Research topics are stochastic optimal control and computational integer programming and their applications in energy economics, water management in agriculture, logistics and epidemiology.

• 5 ongoing PhD projects

https://www.b-tu.de/index.php?id=27930&L=0

02/2018 - 12/2022, 5 years

Maschineller Spracherwerb. Verstärkungslernen minimalistischer Grammatiken

Abstract

We are developing a reinforcement learning algorithm for numeral words of natural languages. Numerals have an obvious connection between syntax and semantics, e.g. "two hundred and fourty-five" means 2 * 100 + 40 + 5. Our algorithm is intended to not only learn words but also to predict their integer meaning.
State-of-the-art algorithms try this by slot-filling of digit, i.e. they learn how to decide that 'two' is the hundred, 'four' is the ten… This requires a high degree of individual supervision for single languages – especially if their number systems are not even decimal. Instead, we use more general constraints of numerals word structure that are derived from Hurford's theory of the Packing Strategy.
Our algorithm forms new words via exchange of subnumerals and their meaning is predicted by linear regression, see figure.
The exchange of subnumerals works by generalizing expressions like "two hundred and fourty-five" and converting them to an output set of words (X,Y) |--> "X hundred and Y".
In order to decide which subnumerals become generalized and which do not, we use general size and divisibility arguments. We represent the numeral systems by a minimalist grammar (MG).
Moreover, we want to investigate the possibility to utilize search engines as supervisors: Running the wrong word “fiveteen” in a search engine will display significantly less search results opposed to a search of the correct word “sixteen”. Based on that, we attempt to find a model to decide whether or not a word is correct.

• Natural numeral grammars can be efficiently represented as Minimalist Grammars
• Those Minimalist Grammars can be efficiently generated based on a learning algorithm that
  • reinforces itself by linear regression and
  • is supervised by big search engines

11/2020–10/2022

Neural Network Extension for RISC-V AI Applications

Abstract

The implementation of AI processes on embedded microcontroller systems poses a particular challenge due to resource constraints. A possible solution is the use of dedicated accelerator boards in hardware. These enable fast and energy-efficient parallel computation of important computational steps to accelerate AI inference by a large factor. In this project, the standards-compliant implementation of the vector extension as part of the RISC-V core EMSA5 will be a key work package. This includes the implementation of the new instruction set, the creation of test benches, the implementation of a compiler toolchain, and the integration of this additional functionality into the AI library TensorFlow Lite. This lays the foundation for the energy-efficient use of AI algorithms on RISC-V-based edge devices for a wide range of use cases.

• Embedded vector extension as hardware implementation for RISC-V
• Compiler support and Tensorflow lite integration
• Demonstrator with accelerated AI calculations and reduced energy consumption

01/2022-12/2022, 12 months

VIT-V: Verfahren der Industrie 4.0 für die Triebwerks-Vorentwicklung
AP4: Prozesse zur Ermittlung von In-Service-Lebensdauern

Abstract

In aircraft engines, critical components such as turbine disks must be monitored with regard to their service life consumption. In this project, the flight data (Engine Health Monitoring (EHM) data) was analyzed with the aid of a neural network to determine the service life consumed for an individual flight. In this way, a condition analysis can be performed and the next maintenance can be better planned (predictive maintenance).

• It has been shown that the neural network delivers good results in many areas and has a time saving of over 90% compared to other methods. For some areas, the influencing factors have yet to be found and integrated. In summary, the method is quite promising, but still needs a lot of work and fine-tuning to be used productively.

07/2018-07/2021, 3 Years

VIT-V: Verfahren der Industrie 4.0 für die Triebwerks-Vorentwicklung
AP2: The artificial engineer: A smart holistic framework for the automated transfer of geometry to analysis models 

Abstract

In aircraft engines, critical components such as turbine disks must be monitored with regard to their service life consumption. In this project, the flight data (Engine Health Monitoring (EHM) data) was analyzed with the aid of a neural network to determine the service life consumed for an individual flight. In this way, a condition analysis can be performed and the next maintenance can be better planned (predictive maintenance).

• Transfer and application of image recognition, shape recognition and segmentation algorithms to the engineering domain
• Development of a smart recognition framework based on shape, function and context arguments
• Development of strategies to digitally mimic engineering logics and decision making in the simulation modelling process.
• Development of a workflow to automatically create large sets of labelled training data in form of simulations in the engineering field in which data is usually quite rare. 
• Application of Graph Neural Networks in the engineering field to learn contextual relations and as conceptual approach to support decision making in the design phase.

07/2018-07/2021, 3 Years

Holistic open-source platform for embedded systems-on-chip

Abstract

The goal of the KI-IOT project is the development and open-source provision of a complete modular platform for the design of system-on-chip (SoC) based on the RISC-V architecture with embedded non-volatile memories (NVM). With the open-source hardware approach for all digital core components as well as licensing models for analog and proprietary function blocks, the entry barriers - especially for SMEs - for deriving application-specific customized processors for diverse applications are to be significantly reduced in the future. In addition to optimized power management for particularly energy-efficient SoCs for use in applications with limited energy budgets, accelerator components for AI algorithms are to be developed. While one component uses classical flash cells, memristive memory cells are used for a second variant.

• Simulation of Al2O3/HfO2 based devices - Development and characterization of Al2O3/HfO2 based devices
• Process integration of Al2O3/HfO2 based devices
• Development of an Incremental Reset and Verify technique showing enhanced variability and reliability features compared with a traditional refresh-based approach.

01.07.2021-30.06.2024

Energy-efficient data processing in the autonomous vehicle using a multiprocessor system and integrated AI accelerators.

Abstract

The focus of the project is the development of suitable sensor and sensor processing components to build suitable infrastructures for Kl-controlled autonomous driving. The objective here is a hierarchically structured approach so that data is not transmitted in aggregate to a central, powerful (and potentially error-prone) computing instance, but is preprocessed already embedded in the respective sensor units. This approach is intended to ensure that the resulting overall computing platform not only provides sufficient computing power, but in particular the required properties in terms of high reliability, availability, robustness and energy consumption.
In this project, IHP will focus on investigating RRAM technology for neuromorphic accelerators. 

• Optimization of the RRAM cells with regard to neuromorphic requirements
• Architecture of the RRAM array with interface optimized for reading in and reading out neuromorphic data
• Design and fabrication of RRAM arrays optimized for AI applications

01.10.2019-30.09.2022

Memristives In-Memory-Computing: Radiation hard Memory for Computing in Space

Abstract

Electronic circuits used in space have to be radiation-hard. Therefore, it is very attractive to use memristive memory technologies in satellites, e.g. for Internet-of-space (IoS) applications, since the information carrier in memristive devices is based on ions and not on charges like in other memory technologies like Flash. Ions are much less sensitive to space radiation than electrons. Therefore, non-volatile memories (NVM) based on ions are well suited for space applications.
In particular in IoS applications this makes sense since we avoid by in-memory computing comparatively long data transfer along wires being very susceptible to radiation if the data is transferred from memory to the processor core. In addition, this requires to make the unavoidable CMOS transistors, we need for accessing the ReRAMs, radiation-hard, too.

• Characterizing the radiation hardness of RRAM devices in the 130 nm technology node and use the results to extend existing device models with this property
• Optimize physical processes to improve the radiation hardness of 1T-1R topology
• Conceive, realize, prototype and evaluate a radiation hard memory
• Comparative analysis of SystemC-AMS versus mainstream analog simulation engines when running the fundamental circuit simulation types of a resistive crossbar array

https://memristec.de

01.09.2020-31.08.2023

Neutronics - From CMOS integrated devices to the behavioural design of memristive circuits

Abstract

The progress on memristive circuits and systems is gained rapidly. Many international groups studying applications in non-volatile memory and enhanced computing technologies; several are now working intensively on hardware artificial neural networks (ANNs) with memristive devices as synaptic elements.
Nevertheless, there is still a huge gap between the physical implementation and the verification of circuits and systems proposed for memristive devices. The first step, required to fill the gap, is the development of analog simulation tools, which are the base for the successful implementation of digital CMOS circuits with memristive elements. New designs and concepts need to be supported up by physical implementation and verification to be reliable. That means, new simulation tools for memristive devices have to address the following issues: device variability, cycling variability, data endurance, data retention as well as device switching speed.

• Establishing a design enabling platform
• Establishing an analog-level and mixel-signal simulation
• Evaluation of memristive modelling and simulation tools

www.crc1461-neurotronics.de

01.01.2021-31.12.2023