Spatial Hyperparameter Optimisation for Full Waveform Inversion Algorithms

This study highlights the critical role of hyperparameter selection in Full-Waveform Inversion (FWI) workflows, particularly in addressing challenges like non-linearity and cycle-skipping in complex geological environments. A case study using data from the North West Shelf, Australia, showcases the impact of various hyperparameters on inversion stability and accuracy. The dataset, incorporating realistic complexities, was analyzed using multiple algorithms, including Adaptive Waveform Inversion (AWI), Reflective Waveform Inversion (RWI), and conventional FWI. Comparative results revealed the limitations of traditional FWI, particularly in resolving deep structures, and the benefits of combined AWI-RWI workflows.

A systematic search of AWI parameters revealed significant variability in outcomes, with over 30% divergence in correlation scores across parameterizations. Spatial variability in parameter performance was observed, underscoring the need for localized parameter optimization. The study employed metaheuristic algorithms like Quantum Particle Swarm Optimization (qPSO) to automate global hyperparameter optimization, leveraging AWS Cloud resources to efficiently process over 50,000 virtual CPUs.

The findings emphasize the need for advanced objective functions and metaheuristic optimization to handle the expanded hyperparameter space, enabling robust and scalable FWI solutions. Future research should integrate these techniques with cost-effective cloud computing to improve inversion accuracy and efficiency in real-world applications.