Oral Presentation Sub22 Conference

The second-best way to solve a problem is sometimes by machine learning. The best way is to understand the physics and use a physics-based machine learning strategy.

I review the current progress in inverting geophysical data by physics-informed machine learning (ML) algorithms which find the optimal geophysical model m that best explains the data d. The inverse solutions by ML and a physics-based algorithm can be combined to give a hybrid ML+physics algorithm denoted as physics-informed ML. The benefit of combining both strategies in parallel is that the ML algorithm is trained to recognize a familiar data pattern, and so encourages a physics-based gradient algorithm to search in the neighboring model space. This parallel strategy is illustrated in Figure 1b. The hope is that this search avoids many local minima that surround the global minima. Another hybrid strategy is sequential where ML is first used to reduce the dimension of the input data, and then a physics-based inversion is used to estimate the optimal model from the reduced data z. In this case, as illustrated in Figure 1a, the simplest data components are input into the iterative physics-based inversion and so leads to a physics-based misfit function with fewer local minima. This is a type of multiscale strategy, where the unimportant data are deliberately avoided and only the most important parts are inverted for. Results from these two strategies and other physics-informed ML methods are presented for both seismic, electromagnetic and potential field data. I also discuss progress in using an ML algorithm that efficiently solves the wave equation.