The Earth Resources Laboratory (ERL) is MIT’s home for geophysical research driven by technological questions. The laboratory is comprised of a dozen faculty members and their groups, active in areas ranging from seismology to geomechanics, rock physics, flows in porous media, and methods of inversion, inference, and uncertainty quantification. As the “information revolution” is shaking up the research enterprise in many fields, ERL is embracing scientific machine learning as a main research objective. New tools lead to new questions, such as:

• Are estimation and prediction still possible when the physical models are too coarse, or contain too much uncertainty, but when data are abundant?

• How can we bridge the “transfer learning” gap from synthetic to real data, or from labeled (rich) to unlabeled (poor) data?

• Uncertainty quantification in machine learning: what level of confidence should we give to the predictions that come from a neural network?

• Is it possible to automate tasks that otherwise require a human’s ability to make generalizations?

Machine learning and artificial intelligence will only succeed in the sciences if their predictive power can outperform that of human-designed physical or statistical models. ERL has a long-term goal to identify the questions in geophysics, broadly understood, where machine learning genuinely extends the reach of traditional predictive models and data processing. You will find examples in this document.