Practical trainings, student assistants and theses
| Offer | All | School practical training | Master theses | Research Assistant | Student practical training | | Diploma theses | Bachelor theses | Student Assistant | Compulsory internship |
|---|---|
| Institute/ Dep. | All | FWGP-I | FWGT-B | FWIO-T | | FWPC | FWU |
| Formatting | Table | |
Predicting Laser Ion Acceleration from Beam Properties Diagnostics using Deep Neural Networks (Id 473)
Master theses / Diploma theses / Student Assistant / Compulsory internship / Volunteer internship
Particle accelerators use very strong electric fields to accelerate charged particles to high kinetic energies. Due to limits on the strength of the acceleration field per unit length (think voltage drop over a capacitor), the acceleration path must be very long (kilometers) to reach very large particle energies. High-power lasers can produce short pulses of extremely intense light, i.e. bunches of electro-magnetic waves with a large amplitude, producing far stronger accelerating fields over a short distance than conventional accelerators can achieve. Hence, high-power lasers can accelerate charged particles to GeV energies over a short distance (millimeters). However, the laser field is much more complex and harder to control than that of a linear accelerator. Hence, obtaining high performance consistently is a big challenge—which scientists at HZDR are working on. To better understand laser particle accelerations (LPA) experiments, we want to learn how parameters of the driving laser correlate with the results of the LPA process, e.g. with the amount (charge) or kinetic energy of the accelerated particles. For this purpose deep neural networks cna be employed to extract any non-linear relations between inputs (e.g. laser spectrum) and outputs from existing data.Tasks
- Familiarize yourself with the modular implementation deep neural networks in PyTorch and an existing code base.
- Plan and run hyperparameter optimization of a hierarchical network architechture, balancing training down-stream tasks and (pre-training) of autoencoders for input representations.
- Experiment with different network architechtures to extract smaller, more expressive latent representations of input data.
- Evaluate and present your results following scientific principles.
- Speed-up the PyTorch-based training code using distributed processing and profile and optimize it to decrease training time.
Department: Laser Particle Acceleration
Contact: Dr. Kelling, Jeffrey, Dr. Pandit, Vedhas Sadanand
Requirements
- Programming knowledge in Python and at least one other programming language
- Experience working with machine learning framework like PyTorch.
- Basic knowledge of machine learning, including deep learning.
- Knowledge of working with a Linux/Unix shell and version control systems (Git)
- Ability to work independently and systematically solve problems
Conditions
- Duration of at least 6 Months
- Option to extend the topic into a Master/Diploma thesis
