Buyu Deng 邓卜瑜

Subsurface Energy Systems · Deep Learning

I am an undergraduate student majoring in Geophysics at the University of Science and Technology of China (USTC), expecting to graduate in June 2027.

At USTC, I conduct research in the Computational Interpretation Group under the supervision of Prof. Xinming Wu. My work focuses on deep learning for seismic data.

I am also an Undergraduate Research Intern at the Stanford Center for Earth Resources Forecasting (SCERF), working with Prof. Tapan Mukerji on geomodeling and forward modeling.

I plan to pursue a Ph.D. and am broadly interested in Artificial Intelligence and Subsurface Energy Systems.

Research Affiliations

Computational Interpretation Group (CIG) University of Science and Technology of China Stanford Center for Earth Resources Forecasting (SCERF) Stanford University

Research Interests

  • Seismic interpretation, geomodeling, forward modeling
  • Deep learning, agentic AI

News

Gave an oral presentation on Geo‑VAE at IMAGE 2026 in Houston.

Joined Stanford SCERF as an Undergraduate Research Intern.

Joined the Computational Interpretation Group at USTC.

Publications & Presentations

  1. Yimin Dou, Xinming Wu, Zhixiang Guo, Hui Gao, and Buyu Deng. Implicit Structural Modeling via Generative Diffusion Frameworks. arXiv preprint arXiv:2606.07165, 2026. [arXiv] [DOI]
  2. Geo‑VAE: A Unified Variational Autoencoder for 3D Geophysical Data Compression and Latent Processing. Oral presentation, International Meeting for Applied Geoscience & Energy (IMAGE 2026), Houston, Texas, August 2026. Full manuscript under revision at Geophysics. [Conference]

Selected Projects

Geo‑VAE

Geo‑VAE adapts the pretrained Wan‑VAE to map seismic, impedance, RGT, fault masks, and field seismic into a compact latent domain. Cached causal resampling treats crossline as a pseudo-temporal sequence, yielding isotropic 8×8×8 downsampling and 31:1 scalar compression for 256³ cubes.

Across 128³–512³ cubes, it maintains competitive, modality-dependent reconstruction fidelity and reduces latency versus fine-tuned Wan‑VAE at larger volumes. The latent interface supports RGT-conditioned seismic generation and complete-volume 768³ latent denoising.

[Code]

Overview of Geo-VAE, showing cached causal 8×8×8 compression, RGT-conditioned seismic generation, and latent denoising
Cached causal compression in Geo‑VAE and its RGT-conditioned seismic generation and complete-volume latent denoising tasks.

All-in-One 3D Seismic Denoising

Degradation-representation learning for robust suppression of random noise, acquisition footprints, and migration artifacts in 3D seismic data. [Code]

PHY-to-BGC Ocean Biogeochemical Emulation

A 2.5D full-image U-Net emulator mapping global physical ocean fields to coupled biogeochemical variables, with evaluation under anomalous ocean-climate conditions. Manuscript in preparation.

GNSS–NISAR Ionospheric Reconstruction

An automated pipeline aligning GNSS TEC-derived ionospheric phase with NISAR dual-frequency InSAR for high-resolution super-resolution research.

Experience & Education

University of Science and Technology of China

B.S. Candidate in Geophysics, School of the Gifted Young

Selected Honors

  • Zhao Jiuzhang Elite Program Scholarship
  • Zhao Jiuzhang Elite Program Scholarship
  • Rose Fund Endeavor Scholarship
  • Freshman Scholarship, USTC