High Performance Computing Group

Bridging the Gap between Scientific Simulations and Experiments with Cycle-Consistent Generative Models

Many large-scale physics experiments, such as ATLAS at the Large Hadron Collider, Deep Underground Neutrino Experiment, and sPHENIX at the Relativistic Heavy Ion Collider, rely on accurate simulations to inform data analysis and derive results. Any systematic biases in the simulations can propagate through the analysis steps and result in systematic uncertainties. These biases may be detected and corrected using heuristics in the conventional analysis workflow, but they may become intractable when machine learning (ML) and artificial intelligence methodologies are applied, which tend to train on the simulations and infer on real data. Here, we are developing a physics-informed ML framework that can bridge the gap between simulations and experiments. This potentially can be realized by applying Generative Adversarial Networks (GANs) in an innovative way. We are constructing a GANs-based Cycle-Consistent Cross-Domain simulation augmentation framework to improve scientific simulations and better represent reality. We also are seeking ways to encode relevant physics conservation laws and symmetries into the framework while applying effective high-performance computing techniques for scalability.

This novel application of cross-domain GANs represents a paradigm shift from using GANs as surrogate models to speed up simulations to reducing and removing systematic biases with the knowledge of real data. CSI is overseeing the project’s development of new ML architectures with our collaborators from Brookhaven’s Physics Department leading the validation and testing of the developed ML models.

Publications

Torbunov, Dmitrii, et al. "Uvcgan: Unet vision transformer cycle-consistent GAN for unpaired image-to-image translation." Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 2023.

Huang, Yi, et al. "Unsupervised Domain Transfer for Science: Exploring Deep Learning Methods for Translation between LArTPC Detector Simulations with Differing Response Models." arXiv preprint arXiv:2304.12858 (2023).

Torbunov, Dmitrii, et al. "Rethinking CycleGAN: Improving Quality of GANs for Unpaired Image-to-Image Translation." arXiv preprint arXiv:2303.16280 (2023).