Funded OA publications

A fabrication strategy for high-efficiency passive three-dimensional microfluidic valves with no mechanical parts fabricated in silicon and glass substrates is presented. 3D diffuser-nozzle valve structures were produced and characterized in their added value in comparison to conventional diffuser-nozzle valve designs with rectangular cross sections. A grayscale lithography approach for 3D photoresist structuring combined with a proportional transfer by reactive ion etching allowed to transfer 3D resist valve designs with high precision into the targeted substrate material. The efficiency with respect to the rectification characteristics or so-called diodicity of the studied valve designs is defined as the ratio of the pressure drops in backward and forward flow directions. The studied valve designs were characterized experimentally as well as numerically based on finite element simulations with overall matching results that demonstrate a significantly improved flow rectification of the 3D valves compared to the corresponding conventional structure. Our novel 3D valve structures show, for instance, even without systematic optimization a measured diodicity of up to 1.5 at low flow rates of only about 10 μl/s.

The Processing-in-Memory (PIM) paradigm promises to accelerate data processing by pushing down computation to memory, reducing the amount of data transfer between memory and CPU, and - in this way - relieving the CPU from processing. Particularly, in in-memory databases memory access becomes a performance bottleneck. Thus, PIM seems to offer an interesting solution for database processing. In this work, we investigate how commercially available PIM technology can be leveraged to accelerate query processing by offloading (parts of) query operators to memory. Furthermore, we show how to address the problem of limited PIM storage capacity by interleaving transfer and computation and present a cost model for the data placement problem.

Evaluating robustness is an important goal in simulation-based analysis. Robustness is achieved when the controllable factors of a system are adjusted in such a way that any possible variance in uncontrollable factors (noise) has minimal impact on the variance of the desired output. The optimization of system robustness using simulation is a dedicated and well-established research direction. However, once a simulation model is available, there is a lot of potential to learn more about the inherent relationships in the system, especially regarding its robustness. Data farming offers the possibility to explore large design spaces using smart experiment design, high performance computing, automated analysis, and interactive visualization. Sophisticated machine learning methods excel at recognizing and modelling the relation between large amounts of simulation input and output data. However, investigating and analyzing this modelled relationship can be very difficult, since most modern machine learning methods like neural networks or random forests are opaque black boxes. Explainable Artificial Intelligence (XAI) can help to peak into this black box, helping us to explore and learn about relations between simulation input and output. In this paper, we introduce a concept for using Data Farming, machine learning and XAI to investigate and understand system robustness of a given simulation model.

Application problems can often not be solved adequately by numerical algorithms as several difficulties might arise at the same time. When developing and improving algorithms which hopefully allow to handle those difficulties in the future, good test instances are required. These can then be used to detect the strengths and weaknesses of different algorithmic approaches. In this paper we present a generator for test instances to evaluate solvers for multiobjective mixed-integer linear and nonlinear optimization problems. Based on test instances for purely continuous and purely integer problems with known efficient solutions and known nondominated points, suitable multiobjective mixed-integer test instances can be generated. The special structure allows to construct instances scalable in the number of variables and objective functions. Moreover, it allows to control the resulting efficient and nondominated sets as well as the number of efficient integer assignments.

The dynamics in a confined turbulent convection flow is dominated by multiple long-lived macroscopic circulation states that are visited subsequently by the system in a Markov-type hopping process. In the present work, we analyze the short transition paths between these subsequent macroscopic system states by a data-driven learning algorithm that extracts the low-dimensional transition manifold and the related new coordinates, which we term collective variables, in the state space of the complex turbulent flow. We therefore transfer and extend concepts for conformation transitions in stochastic microscopic systems, such as in the dynamics of macromolecules, to a deterministic macroscopic flow. Our analysis is based on long-term direct numerical simulation trajectories of turbulent convection in a closed cubic cell at a Prandtl number Pr=0.7 and Rayleigh numbers Ra=10^6 and 10^7 for a time lag of 10^5 convective free-fall time units. The simulations resolve vortices and plumes of all physically relevant scales, resulting in a state space spanned by more than 3.5 million degrees of freedom. The transition dynamics between the large-scale circulation states can be captured by the transition manifold analysis with only two collective variables, which implies a reduction of the data dimension by a factor of more than a million. Our method demonstrates that cessations and subsequent reversals of the large-scale flow are unlikely in the present setup, and thus it paves the way for the development of efficient reduced-order models of the macroscopic complex nonlinear dynamical system.