2022
Differentiable Physics: a learnable Physics solver
Chair for Graphics and Visualization, TU München
- Topic offered by Prof. Thuerey as part of the course Advanced Deep Learning for Physics
- Implemented and compared physics solvers based on a given potential landscape. Task: predict the trajectory of a point.
- Implemented a supervised learning solver: it is purely data-driven. Problem: leads to suboptimal solutions in case of multiple solutions.
- A differentiable physics solver: it is a data-driven learnable solver that also incorporates knowledge of the potential landscape. It is more accurate than the supervised technique.



Include three pictures on gdrive. Maybe change the first picture by including also the upper trajectory. introduce the phyisics solver: it will provide the baseline here. comment on zig zag in pic3.
2021
Optimization of Artificial Neural Networks: Cascade Correlation Algorithm
Vitesco GmbH and Chair of Numerical Mathematics, TU München
- Supervised by Dr. T. Köpple, MSc. G.Gutierrez.
- Building an optimal Artificial Neural Network (ANN), with a minimal number of layers and neurons, that does not overfit the data.
- Goal: develop an ANN able to run on a small engine processor with a limited amount of memory. The ANN predicts temperature of the vehicle’s engine.
- The Cascade Correlation (Cascor) Neural Network is both an architecture and a family of learning algorithms: it begins with a minimal network structure and then trains and adds automatically units, one at a time, optimizing residual correlations.


pictures: prediction (last picture from report), and this pic https://www.google.com/url?sa=i&url=https%3A%2F%2Ftowardsdatascience.com%2Fcascade-correlation-a-forgotten-learning-architecture-a2354a0bec92&psig=AOvVaw22CIi2HetqtBKnFT9qre2W&ust=1665592532396000&source=images&cd=vfe&ved=0CAsQjRxqFwoTCMC17O_N2PoCFQAAAAAdAAAAABAE
Comparison of simple classifiers
Chair of Mathematical Modeling of Biological Systems, TU München
- Topic offered by Prof. Theis as part of the course Statistical Learning.
- Comparison of different classifiers: Linear Support Vector Machine (SVM), Radial Basis Function (RBF) SVM, Decision Tree, Random Forest and AdaBoost.
- This exercise shows the limitations of linear classifiers (Linear SVM) against kernel classifiers (RBF SVM) and exposes the overfitting behavior of single classifiers (Decision Trees) against ensemble methods of bagging (Random Forest) and boosting (AdaBoost).

include image of exercise.
Investigation of Silicon Properties using the Quantum Espresso simulation environment
Associate Professorship Simulation of Nanosystems for Energy Conversion, TU München
- Supervised by M. Rinderle, I. Kouroudis. Collaborated with J. Schwend.
- Performed self-constistent calculations to obtain density of states (DOS) and bandstructure of Silicon.
- Used different types of exchange-correlation functionals to obtain finer results in bandgap calculation.



include pictures of silicon crystal structure, density of states, bandgap.
2019
Study of parameters that influence radiative transfer in the atmosphere
Condensed Matter Physics Group, Università degli Studi di Milano
- Supervised by Prof. Potenza.
- The greatest challenge in climate research is to determine the effect of light when small particles (aerosol) are present in the atmosphere.
- Studied sensibility of diffused irradiance due to presence in the atmosphere of different Antarctic mineral aerosol dust.
- Carried radiative transfer simulations to determine influence of three crucial aerosol parameters: phase function, extinction cross section and single scattering albedo.


image of radiative forcing from IPCC (last part of thesis), and pg. 29 thesis, first figure: relative differences due to different shapes of aerosols influences irradiance, especially at sea level.