Pei Zhang is a computational scientist in the Computational Sciences and Engineering Division at Oak Ridge National Laboratory. Her research focuses on scientific machine learning for multiscale and multiphysics systems, with interests spanning foundation models, graph neural networks, and scalable AI for scientific computing. Her recent work includes the development of MATEY, a multiscale transformer framework for modeling spatiotemporal physical systems. Pei received her PhD in Aerospace Engineering from Purdue University, where her research focused on high-fidelity modeling of turbulent reacting flows.
Presentation Title:
MATEY: A Multiscale Transformer Foundation Model for Diverse Physical Systems
Presentation Abstract:
Foundation models hold promise for modeling multiscale physical systems—central to applications in energy generation, earth sciences, and power and propulsion systems—with a single base model. By learning across systems and datasets, a pretrained model could provide fast predictions and low-cost adapting to downstream tasks.
Developing foundation models for multiscale multiphysics remains challenging. Transformers, despite their remarkable scalability, often struggle to capture fine-scale local features, while extremely high-resolution spatiotemporal data makes finest-scale tokenization impractical. These challenges are further compounded by multiphysics pretraining across heterogeneous datasets with different variables, resolutions, and modalities.
In this talk, Pei Zhang will present MATEY, a multiscale transformer foundation model for diverse physical systems. The talk will highlight recent advances in heterogeneous pretraining across physical systems, resolutions, and modalities. Fine-tuning results across distinct physics and learning tasks will demonstrate MATEY’s adaptability and remaining challenges for foundation modeling of complex physical systems.