Theory and Computation of PDE Team

邓权灵

2026年10月15日 15:00 至 16:00

Online

Zoom 518 868 7656 (BIMSA)

Antarctic sea-ice concentration forecasting must resolve complex spatial structure while tracking dependencies that span months and a strong but changing annual cycle. This talk presents a hybrid convolutional Transformer for monthly forecasts up to one year ahead. A CNN or UNet encoder compresses each field before factorized temporal and spatial self-attention, preserving local structure while making global mixing practical. Two additions expose the seasonal cycle directly to the model: Fourier-based month-aware positional encoding and a learnable cosine bias on temporal attention. Using NSIDC Sea Ice Index data from 1978-2024, the proposed variants outperform statistical baselines and SEAS5 at the reported horizons while remaining competitive with a multi-source UNet despite using sea-ice concentration alone. Ablations show larger gains at the 12-month horizon, and attention maps reveal clearer periodic structure. The talk concludes with limitations and directions for physics-guided forecasting.

Dr. Quanling Deng is an Assistant Professor at the Yau Mathematical Sciences Center (YMSC), Tsinghua University and an Honorary Lecturer at the Australian National University. Before joining YMSC, he held positions at Curtin University (Research Associate, 2016–2020), the University of Wisconsin–Madison (Van Vleck Visiting Assistant Professor, 2020–2022), and the Australian National University (Lecturer, 2022–2025). He got his PhD in Mathematics at the University of Wyoming in 2016. He has a broad interest in Computational and Applied Mathematics and AI. Inspired by Hilbert’s Sixth Problem, he pioneered a particle–continuum multiscale modeling paradigm for sea ice floes and developed the LEMDA framework that unifies multiscale observations. He also introduced SoftFEM, SoftIGA, and high-order boundary penalty techniques, resolving long-standing challenges in finite elements and isogeometric analysis. His research has been published in leading venues, including MCOM, CMAME, JAMES, ICLR, and the SIAM journals, with over 50 peer-reviewed publications to date.