Research
My research develops computational methods at the intersection of Scientific Machine Learning and geophysics — from physics-informed neural operators and generative models to real-time seismic monitoring and edge-deployable AI.
Research areas
Neural operators & physics-informed learning
Neural Operators (FNO, GNO), Physics-Informed Neural Networks (PINNs), and foundational models for scientific data.
Topics: FNO · GNO · PINO · DeepONet · physics constraints
Inverse problems & uncertainty
Diffusion models for parameter estimation, joint velocity–source inversion, and uncertainty quantification.
Topics: DDPM · I2SB · joint inversion · EnKF
Real-time seismology
Edge-deployable AI for automated microseismic monitoring, DAS compression, and rapid event detection and localization.
Topics: lightweight FNO · DAS · microseismic · edge computing
Cross-domain applications
Sustainable geo-energy (geothermal, CO₂ sequestration), continuous subsurface monitoring, and digital twin frameworks.
Topics: geothermal · CO₂ · data assimilation · digital twins
Microseismic monitoring & source localization
Rapid detection and localization of seismic events is critical for geothermal operations, hydraulic fracturing monitoring, and induced seismicity management. My work develops neural-operator surrogates and explainable AI tools for these challenging environments.
Explainable event detection
The Explainable PhaseNet project provides Grad-CAM and component-wise Shapley value tools for interpreting P- and S-wave detection decisions. SHAP-gated inference reduces false alarms under pump noise and low SNR conditions (Abdullin et al., 2026).
Neural-operator localization
Fourier Neural Operator surrogates enable rapid microseismic source localization. My work with Utah FORGE field data demonstrates real-time capability while maintaining localization accuracy competitive with physics-based methods (Abdullin et al., 2025). Extensions include U-PINO for DAS observations (Al-Qadasi et al., 2026).
Generative models for inverse problems
Ill-posed geophysical inverse problems benefit from generative priors that capture complex subsurface distributions. I develop Diffusion model frameworks (I2SB, DDPM) for joint parameter estimation with built-in uncertainty quantification.
Joint source–velocity inversion
A Schrödinger bridge (I2SB) framework translates observed seismograms into jointly estimated source locations and velocity models, providing uncertainty maps without expensive MCMC sampling (Abdullin et al., 2026).
Parameter-efficient transfer learning
Transfer learning with neural operators enables adaptation to new geological settings with limited labeled data, reducing computational cost while maintaining detection performance (Abdullin et al., 2026).
Technical skills
Scientific Deep Learning
Generative Models (DDPM, I2SB), Neural Operators (FNO, GNO), Explainable AI (SHAP, GradCAM), Physics-Informed Machine Learning
Signal Processing & Inverse Theory
Ill-posed inverse problems, Data Assimilation (EnKF), High-fidelity data compression, Time-series analysis, FFT
Computing & Software
PyTorch, CUDA Optimization, HPC environments, Python (NumPy, SciPy, ObsPy), Linux/Unix