WP3 - Soil property predictions and inversion
Seismic inversion is a technique that transforms seismic reflection data into quantitative subsurface properties like impedance, wave velocity, attenuation and density. These properties—integrated with geological interpretation and other seismic attributes that capture variations in amplitude, frequency and texture—may provide input for statistical and ML prediction models of physical properties of the subsurface.
WP3 aims to provide uncertainty estimates from both seismic inversion and subsequent property prediction that are compatible with the overarching BDA framework being developed across the project. This allows uncertainties to be mathematically combined and propagated through to final design parameters. For this, we have two main tasks:
- We are updating existing (acoustic and elastic) inversion schemes to include Bayesian uncertainty quantification, moving beyond deterministic solutions to probabilistic impedance models. This involves reviewing, selecting, and implementing techniques incorporating Bayesian statistics and deep learning to capture the full range of possible subsurface models consistent with seismic data.
Building on previous work by Dujardin et al. (2020) on physics-based machine learning approaches, we're developing Deep Learning Neural Network (DNN) surrogate models trained on existing inversion outputs and optimized for uncertainty retention and computational speed—critical for handling the large datasets typical of offshore wind projects. - We are developing BDA-compatible statistical and ML models for prediction of CPT and engineering design parameters based on geophysical and geotechnical data, including quantitative uncertainty from the last step. Here, we implement and compare multiple uncertainty quantification methods including ensemble approaches, conformal prediction (demonstrated by Griffiths et al. (2024)), and Bayesian machine learning frameworks.
Figure 1 Cross-sections through a subsurface model comparing acoustic impedance derived from synthetic seismic data though classical inversion (left panels), and using a deep learning-based surrogate model (right panels) (Dujardin et al. 2022)
References:
Dujardin, J. R., Sauvin, G., & Vanneste, M. (2020). Acoustic Impedance Inversion of High Resolution Marine Seismic Data with Deep Neural Network. NSG2020 4th Applied Shallow Marine Geophysics Conference, 1–5. https://doi.org/10.3997/2214-4609.202020169
Griffiths, L., Klinkvort, R. T., Sauvin, G., Vanneste, M., Beyer, S., Dujardin, J., & Vardy, M. E. (2024). Predicting Cone Penetration Tests from Seismo-Acoustic Data: Reliable Uncertainty Quantification Using Conformal Prediction. Fifth EAGE Global Energy Transition Conference & Exhibition (GET 2024), 1–5. https://doi.org/10.3997/2214-4609.202421177
