Research
SEEM develops computational geoscience methods that connect field observations, wave physics, and machine learning. Our current portfolio spans seismic-event intelligence, fiber and borehole sensing, physics-informed operators, imaging and inversion, and transparent geological interpretation.
Research portfolio
Seismic monitoring
Detection, characterization, and localization of small seismic events in noisy field environments.
Topics: microseismicity · moonquakes · induced seismicity
Imaging and interpretation
Reconstruction, inversion, explainable AI, and geological modeling for subsurface understanding.
Topics: FWI · reconstruction · xAI · geological sections
DAS & borehole geophysics
Signal enhancement, inversion, and monitoring with distributed acoustic sensing and vertical seismic profiles.
Topics: DAS · VSP · denoising · lookahead inversion
Physics-informed AI
Neural operators and physics-constrained learning for travel times, wavefields, and inverse problems.
Topics: FNO · PINO · DeepONet · wave simulation
Seismic monitoring and event intelligence
SEEM studies how small seismic events can be detected, explained, and located rapidly enough to support scientific analysis and operational monitoring. The work ranges from terrestrial microseismicity and induced events to the sparse, noisy recordings collected during the Apollo missions.
Detection under difficult noise
Current work combines learned detectors with interpretable evidence. The Explainable PhaseNet project uses Grad-CAM and component-wise Shapley values to inspect P- and S-wave decisions and evaluate explanation-guided detection under white and field harmonic noise (Abdullin et al. 2026). Related research explores synthetic low-magnitude earthquake generation for data-limited settings (Alsulami et al. 2025) and Fourier neural operators for detecting moonquakes in waveform and spectrogram representations (Al-Qadasi and Waheed 2026).
Localization and monitoring design
Neural-operator surrogates are being developed for rapid source localization with conventional and DAS observations. The FNO Localization and U-PINO DAS Localization projects use Utah FORGE data to study fast, resolution-aware, and physics-guided location methods (Abdullin et al. 2025; Al-Qadasi et al. 2026). This direction is complemented by research on sensor placement, shallow-borehole signal-to-noise performance, and induced-seismicity bounds (Dia et al. 2024; Kalinichenko et al. 2025; Cao et al. 2024).
Imaging, inversion, and seismic reconstruction
Imaging and inversion convert incomplete or indirect observations into subsurface models. SEEM combines physical constraints, learned priors, operator learning, and carefully designed benchmarks to improve robustness and reproducibility.
Reconstruction of incomplete seismic data
SeisReconNO and UFNO Seismic Reconstruction learn 3D reconstruction operators for random, trace-wise, and regular missing-data patterns. Their U-Net-enhanced Fourier architectures combine global spectral representations with spatial feature recovery and are evaluated across missing ratios and data resolutions (Traversa et al. 2026, 2025).
Full-waveform and geophysical inversion
Current work includes source-independent, physics-constrained full-waveform inversion with real-data applications (Song et al. 2026), the GeoFWI benchmark data set for evaluating deep-learning inversion methods (Li et al. 2026), and physically guided unsupervised inversion of 1-D magnetotelluric models (Goyes-Penafiel et al. 2025). Together with VSP lookahead inversion, these studies emphasize transparent baselines, field evaluation, and reproducible comparisons.
DAS, VSP, and borehole geophysics
Dense fiber-optic and borehole observations offer high spatial resolution but create demanding problems in noise suppression, calibration, inversion, and computational scale. SEEM develops methods that preserve useful seismic information while making these measurements more actionable.
DAS signal enhancement
The Unsupervised DAS Denoising project suppresses strong DAS-VSP background noise without requiring clean labeled targets. Its attention-based deep-image-prior workflow combines patch construction, feature selection, unsupervised denoising, and reconstruction across synthetic and field surveys (Cui et al. 2025, 2024). DAS arrival-time observations also form the measurement basis for physics-informed source localization in U-PINO.
VSP inversion and borehole monitoring
The PyVSP Deep Learning Lookahead project investigates temporal convolutional networks for estimating velocity ahead of the drill bit. Its reproducibility pipeline includes synthetic model generation, blind finite-difference tests, Volve field examples, ablation studies, and bandwidth and phase-sensitivity analyses (Almarzooq et al. 2026). Broader borehole research includes monitoring design for CO2 sequestration and quantitative signal-to-noise assessment (Eisner et al. 2025; Kalinichenko et al. 2025).
Physics-informed learning and neural operators
SEEM uses physical equations as model structure, training constraints, or validation criteria. The goal is not only faster prediction, but learned models that remain useful across grids, velocity models, acquisition geometries, and data resolutions.
Travel-time and eikonal operators
Research in this area includes physics-informed Fourier neural operators, Fourier-DeepONet, and generalized operator-learning solvers for travel-time fields and the eikonal equation (Song et al. 2024; Liu et al. 2026; Archibong et al. 2026). Hard-constrained measurements provide another route to stable neural traveltime tomography (Taufik et al. 2024). These solvers support source localization, tomography, and survey design.
Wave simulation and learned models
SEEM investigates training strategies and architectures that improve accuracy and convergence for geophysical forward problems. Recent work covers enriched DeepONet representations, faster-converging physics-informed neural networks for seismic wave simulation, and high-resolution eikonal tomography (Haghighat et al. 2024; Zou et al. 2025; Cui et al. 2026).
Geological interpretation and reservoir intelligence
SEEM also develops tools that connect computational predictions to geological reasoning. This includes explicit geological representations, interpretable machine learning, and integrated petrophysical and geological constraints.
Conversational geological modeling
GeoSIRR translates geological narratives into a domain-specific language for coordinate-based cross-sections. The workflow validates descriptions and generated topology, visualizes sections, supports conversational refinement, and answers questions using the explicit section representation (Anikiev et al. 2026). This makes the intermediate geology inspectable instead of treating an AI-generated image as the final model.
Reservoir and interpretable models
Research on subsurface characterization combines rock-physics templates, machine learning, petrophysical constraints, and automated model selection for shale sweet-spot delineation and property prediction (Davy et al. 2026; Davy, El-Husseiny, et al. 2024; Davy, Waheed, et al. 2024). Explainability work in seismic detection provides a complementary view of how model evidence can be inspected before predictions are used in scientific or operational decisions (Abdullin et al. 2026).
Open and reproducible research
Public code and traceable research artifacts are part of the group’s research practice. Project pages connect the scientific question to repositories, papers, data sources, and reproduction instructions; the AI Coding Agents Workshop extends these practices into researcher training.