Climate, Life, Earth, and Planets (CLEPS)

Climate:

2005+ Physics-statistics rapidly quantifies tropical cyclone-induced risk in a changing climate.
2017+ Nonparametric tail bounds enable robust risk-informed sustainable decision-making.
2020+ Neural Physics downscales wind, rain, and flood hazards to very high resolutions.
2023+ Machine Learning accelerates the search for storm extremes in a changing climate.
 

A major theme in our work advances ML-Physics and Statistical-Physical methods to rapidly and accurately quantify wind, rain, and flood hazards in a changing climate, coupling with vulnerability and exposure of regions and populations to quantify the risk from extremes. We use the quantified time-dependent risk to solve decision problems to improve sustainable outcomes. Currently, we are studying the south-west Bangladesh, Broward County, Boston, Puerto Rico, Cape Town, and Bangladesh, but our approach is not limited to these. Visit http://crises.mit.edu to learn more!

Life:

1996+ Using Machine Learning and Computer Vision for Conservation.

In continuous development and research since 2002, our Sloop system was the first system to use "generic image features" and their representations with an image retrieval paradigm coupling with humans and crowd-sourcing to identify individual animals from photographs. Our systems today also incorporate deep learning and show amazing performace, learning to improve recall from relevance feedback by asking humans fewer and better questions after only a few iterations of a co-active learning strategy. Visit http://sloop.mit.edu to learn more.

Earth:

2021+ Machine Intelligence Improves Global Seismic Monitoring.
2022+ Neuro-Physical Inverters help discover Geothermal Energy.

We are developing new informative optimization and inference techniques to locate, associate and estimate earthquakes from observations at seismic stations across the world. A key benefit of our approch is to identify the number of events and their location uncertainties much faster than purely sampling-based approaches to inference. 

Planets:

2019+ Using ML-Physics to detect and characterize exoplanets.

Using a counter-factual approach, we learn what the photometric environment without a transit would look like from the vast amounts of unlabeled data, then train detectors using physics and the modest number of labeled examples for what a transiting exoplanet light curve might look like. The results are super fun! We are also interested in analyzing spectral signatures for signs of life.

 

The themes in our group and beyond are explored through CLEPS workshops, see CLEPS22 for the most recent workshop.