Zoom Link

The KIT Graduate School Computational and Data Science (KCDS) at KIT Center MathSEE proudly presents: KCDS Talks, a monthly series of short lectures from basic knowledge to trending topics in computational and data science.

In July, KCDS Fellows present their research in:

1. Neural Nets for Solving Economic Models (Lukas Frank, ECON)
2. Data driven model for weather forecasting (Deifilia To, SCC/IMKTRO)
3. Reconstruction of Particle Position and Size in Dispersed Multiphase Flows using Deep Learning and Physics-Based Optimisation (Christian Sax, ISTM/IANM)

1. Neural Nets for Solving Economic Models (Lukas Frank, ECON)

Solving rich economic models globally often requires to solve a high-dimensional functional equation. With classical grid-based methods, the curse of dimensionality limits the number of model features to d ≈ 20. Recent advances leverage the abilities of neural nets to mitigate the curse of dimensionality, bringing more realistic models in reach. However, neural nets pose new challenges such as mediocre accuracy and fragile convergence behavior. I show how to solve high-dimensional economic models with neural nets and how to cure some of the most salient issues.

2. Data driven model for weather forecasting (Deifilia To, SCC/IMKTRO)

Traditional methods for weather forecasting are based on the solution of physical conservation equations that are grounded in theory. In contrast, current machine learning methods learn only through data. Machine learning methods can now create better forecasts than traditional methods - but their success is not well understood. I replicate and study one of the most well-known models, Pangu-Weather, and propose improvements in the architecture that could lead to more efficient training and accurate weather forecasts.

3. Reconstruction of Particle Position and Size in Dispersed Multiphase Flows using Deep Learning and Physics-Based Optimisation (Christian Sax, ISTM/IANM)

Dispersed multiphase flows play an important role in a multitude of environmental and industrial applications, such as spray, mist, cavitation and boiling. A novel diagnostic tool is developed for the investigation of such flows from single camera images. The approach combines deep learning for image segmentation and classification with the optimization of a non-linear functional incorporating a model of the scattering process.

If you are a master student, a doctoral researcher, a senior researcher or just interested in the topics - join us!

(for free and without registration)