This project's goal is to create neural networks that help design new experiments for the BESSY electron accelerator by training the networks on simulation data. In a first step, the networks are trained and then used to predict the experiment outcome based on input parameters. After that, we attempt to predict the necessary input parameters for a desired outcome. Finally, we try to optimize the networks to function with noisy data.
Project Data
The training data for the neural network are simulations provided by Peter Feuer Forson.
Usage
This code has been tested and developed using PyCharm and Visual Studio Code and Python 3.9. Set hyperparameters and run the code to start training a neural network with the set parameters. If there is already a fully trained model with the set parameters in the "Modelle" folder, the program will evaluate that model. If there is a partly trained model, it will use that as a checkpoint and continue training it. The program saves models every 10 Epochs. It is recommended to train the models on a GPU with CUDA, as this is much faster than training them on a CPU.
Contributors
- Lionel Kress
- Laurids Radtke
- Martin Kühn
- Qiqi Chen
- Michal Brus
- Clemens Paul