Originally posted to LinkedIn, I moved this thread here to give hard links and allow expansion of the original post beyond LinkedIn length limits.
——————————-
I am looking for examples of geoscience problems solved by machine learning. Before you say there are hundreds in the literature, here are the requirements:
——————————-
I am looking for examples of geoscience problems solved by machine learning. Before you say there are hundreds in the literature, here are the requirements:
1. It is an important problem in either a scientific, society or business sense
2. It is a problem that was unsolved until solved by machine learning
3. Results have been independently verified
4. Code and test data are public (free and available)
Items 3 and 4 are not stipulated because I am an academic do-gooder, they go to the heart of reproducible research that holds down the hype on any new technology.
#machinelearning #geophysics #reproducible
8/14/19
- re: Werner Heigl comment
Physics-Guided Neural Networks (PGNNs)
-Feb 2018 arXiv paper
8/15
-GitHub site on physics based deep learning. Several examples here that include paper, code, and some even have test data. Not geoscience oriented, but meeting other criteria in my original post.
-This paper generates model data from neural net without solving PDE, just training on thousands of parameter models and associated expensively-simulated FD model data. “[the NN] outperforms the FD model across all possible grid sizes and the computational performance gain improves as the grid size is increased” (10^3 speed up @ 512x512). Could reverse time migration or full waveform inversion be 1000+ times faster?
https://arxiv.org/pdf/1709.02432.pdf
https://arxiv.org/pdf/1709.02432.pdf
