Internship: Graduate-Level Research in Industrial Projects for Students

Are you a master or PhD-student and interested in an internship in Berlin during summer this year?

Then the Graduate-Level Research in Industrial Projects for Students (GRIPS) might be the right project for you

Located at the Zuse-Institute Berlin (Research Campus MODAL), we offer graduate students in mathematics and related disciplines the opportunity to work on industry-sponsored research problems. (See list of projects below.)

The program will run from June 24 - August 16, 2024.

G-RIPS is a joint activity of ZIB (Research Campus MODAL) and IPAM (Institute for Pure & Applied Mathematics (at) UCLA).

Students from the U.S. and Europe will work for 8 weeks on cross-cultural teams on three research problems designed by the industrial sponsor. The projects will be of serious interest to the sponsor and will offer a stimulating challenge to students; most will involve both analytic and computational work.

Eligible applicants include master and PhD students from the areas of mathematics, computer or natural sciences who are currently enrolled at a European university. Due to the large number of applications we typically receive, we do not accept applications from previous RIPS or G-RIPS students.

Please submit the following documents :

1. motivation letter (please state which project you are interested in),
2. resume / short CV,
3. an academic record or transcript (can be unofficial).

Costs for traveling and accommodation in Berlin are covered.

This year, we offer the following three projects.

\=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-

\=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-

Project 1: Flow instabilities in wind energy systems

Sponsor: I2DAMO (Intelligent Integrated Data Analysis and Mathematical Optimization) GmbH

Project Description:

Motivation: Integrating renewable energy sources,
such as wind power, is critical for sustainable energy production. Wind
farms comprising multiple turbines face challenges related to wake
interactions that can significantly impact overall efficiency. Flow
instabilities, e.g., the Kelvin-Helmholtz instability (KHI), a
hydrodynamic phenomenon that occurs when two fluids of different
densities flow past each other with different velocities, may play a
crucial role in shaping wind farm wakes. The instability of wind flows
around wind turbines may form vortices and thus reduce the efficiency of
turbines. Understanding and mitigating these instabilities is essential
for optimizing wind farm layouts and enhancing energy extraction.

Methodology:

• Using CFD Simulations to model the flow around a single turbine and a
  multi-turbine wind farm via solving the Navier-Stokes equations
• Analyzing the wake characteristics, including the formation and
  evolution of vortices via identifying the locations where flow
  instabilities occur and their influence on the wind farm power output.
• Analyze the effects on turbulence levels, wake dynamics, and downstream turbine performance.
• Determining the optimal turbine spacing and layout to minimize
  instability modes-induced losses and maximize wind farm efficiency,
  e.g., adjusting the distance between turbines
• Identifying the stability condition of the flow around the turbine to prevent the vortex structures (instability)

Expected Outcomes:The project could lead to the development of new design guidelines for wind turbines and wind farm layout by mitigating instability modes-induced losses and optimizing turbine layout, thus improving the overall performance and economic viability of wind energy systems. By minimizing wake-induced
losses and increasing wind farm efficiency, the project can reduce the environmental impact of wind energy production.

Requirements for the Applicants:

• Strong programming skills in C/C++ or Rust for numerical calculations, data analysis, and plotting
• Good knowledge of computational tools, such as CFD software or mathematical modeling and numerical methods (e.g., FEM or VOF)
• Background in fluid dynamics, aerodynamics, or a related field
• Excellent communication skills for reporting findings in a written document and oral presentation
• Ability to work independently and as part of a team to achieve
  project goals, an interest in wind energy research, and a passion for
  learning new concepts

\=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-

\=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-

Project 2: Machine learning for optimizing numerical settings of finite-element based optics simulations

Sponsor: JCMwave GmbH

Project Description: Photonics is a key enabling technology of the 21st century. Data-based methods play an essential role in the development and application of photonic and quantum technologies. In areas such as sensor technology and quality control,
the evaluation of data enables decision-making and quantification processes. In addition to model-based simulations of measurement processes, artificial intelligence (AI) methods are becoming increasingly important here. The MODAL NanoLab is developing methods for efficient, error-controlled and self-adaptive simulation of
light-matter interactions in optical nanostructures.Efficient tuning of accuracy settings of numerical simulations is of crucial importance for the fidelity and interpretability of the corresponding simulation results. For finite-element based simulations, these settings relate to mesh quality, mesh density, polynomial degree of the finite-element
ansatz functions, and other parameters. The goal of this GRIPS project is to investigate ML methods for finding optimal parameter settings. This should then replace the laborious task of manually adjusting accuracy parameters. In the proposed MODAL GRIPS2024 project, participants will get hands-on experience with the commercial programme package JCMsuite, containing a finite-element solver an optimizer based
on a Bayesian formalism. We will use machine learning to find the optimal JCMsuite execution times for the problem of practical interest while treating the desired precision of the final numerical result as an outcome constraint during optimization runs. In this way, we will completely automatize the search process for the optimal finite-element accuracy settings.

Requirements:

• Experience in Python
• Basic understanding of electromagnetic theory and finite elements
• Basic understanding of optimization algorithms
• Interest in working with state of the art numerical tools

\=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-

\=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-

Project 3: Cryo-electron tomography and deep learning for analyzing sub-cellular structure

Sponsor: Thermo Fisher Scientific

Project Description: Cryo-electron tomography (CET) is the only imaging technique that allows observing three-dimensional (3D) information of subcellular structures directly in their native context in a near-atomic resolution. Due to several physical limitations of the acquisition process, multiple instances of the same protein need to be identified and averaged to reveal the structure, a process called sub-tomogram
averaging (STA). However, identification as well as location and orientation approximation of single macromolecules in tomography data using currently available software is still the bottleneck. This is due to a substantial amount of user interaction required for annotating the data being a time-consuming and error prone process. Recently, particle picking in crowded cell environments of CET images was improved using automated methods like deep learning (DL). DL, on the other hand, requires training with large amounts of annotated data which in the case of STA imposes an insurmountable obstacle. As a result, some challenges like particle orientation approximation have not been addressed. One potential solution to the annotation problem is the use of simulated tomograms bringing the annotations for free. However, the discrepancy between synthetic and real sample distribution, namely the domain gap, significantly reduces the performance of models trained on synthetic
data with experimental data. To this end, in this project, we will investigate DL-based models that predict particle orientation from synthetic data. While such models have a successful performance on unseen synthetic test data, results are still poor on experimental data. Therefore, we will investigate how fine-tuning the embedding feature space learned by a pretrained model can be done. For this, we will explore how the transfer of knowledge from simulated to experimental data can be achieved without retraining the whole model from scratch or requiring large amounts of annotated experimental data.

Requirements of applicants:

• Experience with deep learning frameworks like Keras and Pytorch
• Some knowledge of image processing and computer vision
• Ideally, some knowledge of 3D image data and visualization
• Interest in teamwork