UROP Openings

Have a UROP opening you would like to submit?

Please fill out the form.

Submit your UROP opening

Artificial Intelligence for Scientific Discovery and Understanding


Term:

Spring

Department:

8: Physics

Faculty Supervisor:

Marin Soljacic

Faculty email:

soljacic@mit.edu

Apply by:

2/3/20

Contact:

Please send CV and transcript to Samuel Kim at samkim@mit.edu

Project Description

In the last decade, we have witnessed enormous progress in applications of artificial intelligence (AI) for a wide variety of different tasks, including natural language processing, image recognition, self-driving cars, and playing games. We want to investigate how some of these recent AI techniques can be used for scientific discovery. We want to move away from pure data-science approaches since many current deep learning techniques lack interpretability or generalizability (extrapolation outside of the training data set). Rather, we want to more closely integrate physics and machine learning. The main goal of this project is to develop new AI tools to support scientific discovery in broad areas of science and engineering. We will design deep learning architectures that can extract interpretable parameters and derive generalizable laws of nature so that we can automate optimization of experiments, designs, and data analysis techniques. Our datasets are generated from a variety of systems in physics and engineering, including Maxwell’s equations (electrodynamics), diffusion equation, Navier-Stokes equations (fluid dynamics), and Schrodinger’s equation (quantum mechanics). Each undergraduate student will have their own project supervised by a graduate student, and will ideally transition to leading the project and come up with new ideas. Students will learn about a variety of modern deep learning techniques.

Pre-requisites

Comfort with Python and exposure to machine learning is required. Any major of any year is welcome. Prefer experience with Tensorflow/Pytorch. Exposure to deep learning, optimization, differential equations, numerical simulation, and/or Bayesian statistics is helpful, but not required.