GeoBrainArtificial Intelligence × Exploration Geophysics 中文
HomeStarter GuideResearchPeoplePublicationsProjectsNews 中文

Research

From seismic records to general-purpose geophysical intelligence.

01

Seismic Data Processing

Making incomplete and noisy observations usable again

刘Qi Liu成Zhitong Cheng高Tianxiang Gao胡Peng Hu张Chen Zhang李Shirui Li高Wei Gao盛Wenzhe Sheng程Yuying Cheng马Jianwei Ma

We study the full chain from missing-trace reconstruction to first-arrival picking under irregular acquisition, random noise and coherent interference. Generative models, self-supervised learning, implicit representations and geometry-aware networks are combined with seismic structure, while SeismicBench records datasets, metrics and reproducible experiments.

Methods and workstreams

Diffusion models and generative interpolationRandom and coherent noise attenuationSelf-, weakly and unsupervised learningLow-rank, tight-frame and sparse representationsGeometry-aware first-arrival pickingMultiples attenuation and unified benchmarking

Core questions

  1. How can event continuity be preserved under large and contiguous trace gaps?
  2. How can models trained on limited data transfer to new surveys and acquisition geometries?
  3. How should processing methods be compared with unified and traceable protocols?

Related verified work

04
02

Geophysical Inversion

Recovering subsurface properties from observations

马Jianwei Ma刘Qi Liu

For velocity-model building, full-waveform inversion and multiphysics problems, wave equations, sensing priors and deep networks are placed in a common optimization framework. Work spans end-to-end inversion, physics-informed generative learning, neural implicit representations and parameter-efficient adaptation, with attention to missing low frequencies, initial-model dependence, 3D cost and uncertainty.

Methods and workstreams

Full-waveform inversionDeep velocity-model buildingPhysics-informed generative learningSensing-prior constraintsNeural implicit representationsLoRA and multimodal parameter-efficient adaptation

Core questions

  1. How can cycle skipping be reduced without low frequencies or accurate initial models?
  2. How should physical constraints enter data, network architectures and objectives?
  3. How can pretrained representations adapt to industrial 3D data with few labels?

Related verified work

04
03

Artificial Intelligence Algorithms

Robust, transferable and interpretable learning

成Zhitong Cheng高Tianxiang Gao胡Peng Hu葛Yulong Ge马Jianwei Ma

We derive broadly useful AI methods from geophysical problems, including domain adaptation, noise-robust losses, personalized federated learning, generative modeling and efficient 3D vision. The goal is to explain why models transfer, when they fail and how they can be deployed reliably under limited data and compute.

Methods and workstreams

Feature-space domain adaptationNoise-robust loss functionsPersonalized federated learningDiffusion and generative modelingTransformers and neural operatorsMulti-view 3D reconstruction

Core questions

  1. Does domain shift arise from feature degradation or decision-boundary misalignment?
  2. How can robustness remain analyzable under label noise and client heterogeneity?
  3. How can geometry and physical symmetries be encoded in efficient networks?

Related verified work

04
04

Foundation Models for Geophysics

From task-specific networks to adaptable representations

刘Qi Liu马Jianwei Ma成Zhitong Cheng胡Peng Hu葛Yulong Ge

We study scalable pretraining and cross-modal alignment across seismic, gravity, magnetic, electromagnetic and textual knowledge, building foundation models that adapt through prompts, lightweight decoders or low-rank updates. The program also covers data governance, physical consistency, generalization evaluation, trustworthiness and open benchmarks.

Methods and workstreams

Scalable self-supervised pretrainingMultimodal alignment and prompt enginesParameter-efficient fine-tuningCross-survey, cross-modal and cross-task generalizationPhysical consistency and trustworthy evaluationOpen data, models and benchmarks

Core questions

  1. Which pretraining objectives produce representations transferable across modalities?
  2. How should true generalization to unseen surveys and tasks be evaluated?
  3. How can foundation-model outputs obey physics and retain complete provenance?

Related verified work

04
05

Geophysical Agents

Intent-driven, verifiable scientific workflows

刘Qi Liu马Jianwei Ma

We explore geophysical agents that use language models to plan experiments, call specialist software and organize evidence. The first public case exposes SPECFEM 2D, 3D Cartesian and 3D Globe workflows as MCP tools, supporting automated and human-in-the-loop execution from parameter generation and meshing to solving and visualization; future work extends this pattern to processing, imaging, inversion and interpretation.

Methods and workstreams

Model Context Protocol (MCP) toolchainsSeismic-modeling planning and executionScientific software and data orchestrationHuman-in-the-loop parameter and result reviewError diagnosis, recovery and experiment provenanceReproducible agentic geophysics workflows

Core questions

  1. How can agents translate scientific intent into auditable tool calls?
  2. Which decisions must remain subject to geophysicist review and intervention?
  3. How should correctness, reproducibility and recovery be evaluated for automated workflows?

Public tools and demos

First Agent Template · Hugging Face SpaceA public interactive agent prototype Space by Qi Liu.

Related verified work

02