The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Deep Learning for Earthquake Aftershock Patterns with Phoebe DeVries & Brendan Meade - #311
Oct 25, 2019
Phoebe DeVries, a postdoctoral fellow at Harvard, and Brendan Meade, a professor there, delve into the groundbreaking fusion of deep learning and earthquake prediction. They discuss how machine learning can analyze GPS data to forecast aftershock patterns. The conversation highlights the shift to neural networks for faster modeling, revealing surprising insights into seismic activities. Their collaborative efforts aim to create advanced computational tools that integrate tectonic dynamics, enhancing the accuracy of aftershock forecasts and understanding earthquake mechanisms.
AI Snips
Chapters
Transcript
Episode notes
HPC to ML for Earthquakes
- Phoebe DeVries's HPC earthquake modeling took 2.5 million CPU hours.
- This led them to train a neural network to emulate the HPC code, resulting in a 500x speed increase.
Physics-Based Regularization and Insights
- The neural network, trained on stress changes and aftershock locations, rediscovered a known physical concept.
- It identified the von Mises yield criterion, used to explain transitions from elastic to plastic behavior, as a key factor.
Data Assembly Challenges
- The most challenging aspect was assembling the dataset, not building the neural network.
- This involved processing complex, inconsistently formatted slip distribution data from various sources.