Smart Seismic Inversion via Federated Deep Learning Frameworks

Thesis Abstract

The study addresses the persistent challenge of obtaining high-fidelity seismic inversions under data privacy and heterogeneity constraints across multiple operators and jurisdictions. Traditional centralized seismic inversion frameworks require pooling large volumes of raw seismic and well-log data, raising concerns about data ownership, security, and regulatory compliance. This research investigates a federated deep learning approach to seismic inversion that preserves data locality while enabling collaborative model training across distributed datasets. The aim is to develop a robust, privacy-preserving inversion framework that integrates physics-informed neural networks with federated learning to produce accurate subsurface property estimates (e.g., acoustic impedance, shear impedance, and velocity models) from seismic data without sharing raw traces. Specific objectives include (1) designing a federated learning architecture that couples generative deep networks with physics-based forward models to enforce geophysical plausibility; (2) evaluating data heterogeneity and communication-efficient training strategies on multi-operator datasets; (3) quantifying improvements in inversion accuracy against baseline centralized and distributed approaches; (4) assessing the robustness of the framework to partial data availability and label noise; and (5) establishing a set of operational guidelines for deployment in industry environments. The methodology employs a mixed-methods research design combining quantitative algorithmic evaluation with qualitative assessments of operational feasibility. The population comprises three real-world seismic datasets from distinct oil-and-gas basins, each containing 2D seismic lines, well-log measurements, and available impedance models. A stratified sampling approach yields 40 seismic lines per basin, totaling 120 lines, partitioned for training, validation, and federation; data never leaves the local site, in line with privacy-by-design principles. Data collection instruments include standardized seismic amplitudes, synthetic forward-model runs for ground-truth generation, and metadata descriptors (acquisition parameters, noise levels, and well log coverage). The core method of analysis centers on a federated learning framework that integrates a conditional variational autoencoder (CVAE) for latent space representation with a physics-informed neural network (PINN) that enforces the acoustic impedance–velocity–density relationships via a differentiable forward model. Training employs FedAvg with client-specific learning rate schedules, gradient clipping, and differential privacy mechanisms to mitigate information leakage. Model validation uses k-fold cross-validation across basins and Monte Carlo simulations to assess uncertainty. For performance benchmarks, the study applies root-mean-square error (RMSE), structural similarity index measure (SSIM) on inverted impedance profiles, Nash–Sutcliffe efficiency (NSE) on derived velocity fields, and Bland–Altman plots to evaluate bias across datasets. Additional analytical techniques include ablation studies to isolate the contributions of physics-informed regularization and federated aggregation, as well as regression analyses to relate inversion accuracy to data diversity and noise characteristics. Expected findings indicate that the federated physics-informed framework will achieve superior inversion accuracy compared with centralized baselines under equivalent data volumes, with RMSE reductions of 12–18% and SSIM improvements of 0.05–0.12 across test lines. The approach is anticipated to demonstrate enhanced robustness to variable acquisition parameters and partial data coverage, preserving geophysical consistency through PINN constraints. Uncertainty quantification via Bayesian-inspired federated ensembles is expected to reveal narrower credible intervals for impedance estimates compared with non-federated methods. The study anticipates discovering a quantifiable relationship between data heterogeneity and convergence rate, informing optimal client selection and communication scheduling. The contribution to knowledge lies in, first, validating a privacy-preserving, federated learning paradigm for seismic inversion that preserves data confidentiality while delivering high-quality subsurface models; second, advancing the integration of physics-informed deep learning within distributed training to enforce geophysical realism; and third, providing a pragmatic deployment blueprint for industry-scale adoption, including guidelines on data preparation, privacy settings, model governance, and performance monitoring. The main conclusion is that federated deep learning frameworks, when augmented with physics-informed regularization, can deliver reliable seismic inversions in data-sharing constrained environments. Recommendations include extending the framework to 3D seismic data, exploring transfer learning across basins to further reduce data requirements, and engaging with regulatory bodies to codify privacy standards for cooperative subsurface modeling.