Lorentz-Equivariant Geometric Algebra Transformers for High-Energy Physics
Authors: Jonas Spinner, Victor Breso, Pim de Haan, Tilman Plehn, Jesse Thaler, Johann Brehmer
NeurIPS 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We now demonstrate L-GATr in three applications. Each addresses a different problem in the data-analysis pipeline sketched in Fig. 1. 4.1 Surrogates for QFT amplitudes 4.2 Top tagging 4.3 Generative modelling |
| Researcher Affiliation | Collaboration | Jonas Spinner Heidelberg University EMAIL Victor Bresó Heidelberg University EMAIL Pim de Haan Qualcomm AI Research Tilman Plehn Heidelberg University Jesse Thaler MIT / IAIFI Johann Brehmer Qualcomm AI Research |
| Pseudocode | No | The paper describes the architecture and layers mathematically and textually but does not provide a formal pseudocode block or algorithm. |
| Open Source Code | Yes | Our implementation of L-GATr is available at https://github.com/heidelberg-hepml/ lorentz-gatr. |
| Open Datasets | Yes | We use the reference top quark tagging dataset by Kasieczka et al. [49, 50].12 The data samples are structured as point clouds, with each event simulating a measurement by the ATLAS experiment at detector level. ... 12Available at https://zenodo.org/records/2603256 under a CC-BY 4.0 license. |
| Dataset Splits | Yes | Each dataset consists of 4 105 samples for training, 105 for validation, and 5 105 for testing. ... The dataset consists of 1.2 106 events for training and 4 105 each for validation and testing. ... On each dataset, 1% of the events are set aside as validation and test split. |
| Hardware Specification | Yes | Our measurements are performed with datasets made up by a single sample and all models are run on an H100 GPU. |
| Software Dependencies | Yes | The t t + n jets, n = 0...4 dataset is simulated with the Mad Graph 3.5.1 event generation toolchain, consisting of Mad Event [4] for the underlying hard process, Pythia 8 [72] for the parton shower, Delphes 3 [35] for a fast detector simulation, and the anti-k T jet reconstruction algorithm [21] with R = 0.4 as implemented in FASTJET [22]. |
| Experiment Setup | Yes | All models are trained by minimizing a mean squared error (MSE) loss on the preprocessed amplitude targets and by making use of the Adam optimizer. We use a batch size of 256 and a fixed learning rate of 10 4 for all baselines. |