Hello, I am a postdoctoral associate in the MIT Department of Earth, Atmospheric, and Planetary Sciences. My research is tackling climate change with machine learning, little-by-little, together with Prof. Raffaele Ferrari and Prof. Noelle Selin.
I am concerned that running a high-resolution (1km) climate model can take multiple weeks on the world's largest supercomputers; consuming the same electricity a coal power plant would generate in one hour. To overcome the computational complexity, we are reshaping machine learning models into fast copies, or 'surrogates', of climate models. The core difficulty is to ensure physical-consistency in the surrogates, such that policy- or decision-makers can trust the machine learning surrogates.
My research has won grants in the big tech (IBM), government (NSF), nonprofit (CCAI), and space (EAS) sectors. I have been teaching at the Caltech Computer Vision for Ecology Summer Workshops, advised teams at the NASA/SETI Frontier Development Lab, and interned with IBM Future of Climate and John Deere-BRT. I have earned a Ph.D. with Prof. Dava Newman from MIT AeroAstro with a major/minor in Machine Learning and Earth System Modeling, earned an M.Sc. with Prof. Jon How from MIT in safe and robust deep reinforcement learning, and a B.Sc. with Emmanuel Dean from TUM in Engineering Science. I also windsurf poorly.
Click for linkedin , google scholar , our MIT BC3 group , or twitter .
Email: lutjens [at] mit [dot] edu
[Last updated Aug. 22nd 2024]
Emulators are approximations of climate models that can quickly compute climate forecasts when running the full climate model is too expensive. In our latest paper, we examine how different emulation techniques can be compared with each other. We show that a simple linear regression-based emulator can forecast local temperatures and rainfall more accurately than a complex machine learning-based emulator on a commonly used benchmark dataset. It is surprising that linear regression is better for local rainfall, which is expected to be more accurately emulated by nonlinear techniques. We identify that noise from natural variations in climate, called internal variability, is one reason for the comparatively good performance of linear regression on local rainfall. This implies that addressing internal variability is necessary for assessing the performance of climate emulators. Thus, we assemble a benchmark dataset with reduced internal variability and, using it, show that a deep learning-based emulator can be more accurate for emulating local rainfall, while linear regression continues to be more accurate for temperature.
Click for presentation , the preprint , and the code .
In practice, climate models are run at 50-250km resolution per grid cell due to computational constraints. However, some dynamics such as cloud formation requires solving equations at a higher grid resolution. Because that is too expensive, the influence of high- onto low-resolution dynamics is captured in physics- or function-based approximation terms. In our work titled Multiscale Neural Operator, we have shown that neural operators can learn these parametrization terms from very high-resolution simulations while offering a flexible grid representation and cutting compute from quadratic to quasilinear in comparison to fully-resolved solver. Further, there is a concerted effort, called ClimSim, to benchmark the best ML models for parametrizations.
Click for presentation , and the preprint .
Downscaling methods are necessary to map low-resolution climate projections onto resolutions that are relevant for local decision making (1km or less). In this project, we are benchmarking data-driven methods for downscaling surface meltwater in Greenland. We are aiming to create the highest resolution maps of surface meltwater in Greenland. This is work in progress.
Click for an abstract .
I am very grateful for working together, advising, and learning from students:
Ana Mata-Payerro
Lea M Hadzic
Matthew Kearney
Vivian Trinh
Salva Rühling Cachay
Til Widmann
Rupa Kurinchi-Vendhan
Thomas Huber
Ernest Pokropek
Gyri Reiersen
Kenza Amara
Simona Santamaria
Kyle Morgenstein
Rishi Sundaresan
Students of the 2022 CV4E Workshop
Students of the 2023 CV4E Workshop
Simulating coastal floods at high-resolution is computationally too expensive for real-time inference, uncertainty quantification, or low-income countries. We are leveraging neural operators, a new physics-infused deep learning technique, to learn a coastal flood model that is magnitudes faster.
Click for presentation and paper.
As climate change increases the intensity of natural disasters, society needs better tools for climate risk communication. We are creating a "Google Earth of the future"; a global visualization of how climate change will shape our landscape. On our path, we proposed the first deep learning pipeline to ensure physical-consistency in synthetic satellite imagery and are publishing a dataset on 25k labelled satellite images of coastal floods and melting Arctic sea ice. Explore our results at climate-viz.github.io.
Click for presentation , paper , demo , and code .
Small-holder forest conservation projects are often excluded from global carbon markets, because monitoring the forest carbon is too expensive or inaccurate. We create deep learning-based forest inventories from drone and smartphone imagery, collected by locals, to cheaply and accurately estimate the tree-sequestered carbon. The transparent carbon monitoring allows local landowners to participate in the global carbon markets and subsidize their livelihood through maintaining and rebuilding rainforest. Learn more at forestbench.org.
Click for presentation , AAAI '22 paper , and NeurIPS '19 paper .
Deep learning-based models have recently outperformed state-of-the-art seasonal forecasting models, such as for predicting El Niñno-Southern Oscillation. However, current deep learning models are based on convolutional neural networks which are difficult to interpret and can fail to model large-scale atmospheric patterns. We propose the first application of graph neural networks to long range forecasting, because they can better capture spatially distant dependencies. We show that our model, Graphiño, outperforms state-of-the-art machine learning-based models for forecasts up to six months ahead and is more interpretable.
Click for paper .
Deep Neural networks, used in commercially available driver assistance systems, can fail on hardly detectable adversaries, for example, white stickers on the road. This work has invented an add-on, real-time defense algorithm to certify the robustness of Deep Reinforcement Learning algorithms to such adversaries. Patent pending.
MIT News, Algorithm helps artificial intelligence systems dodge “adversarial” inputsDeep neural networks can fail overconfidently on novel observations, for example, an uncollaborative pedestrian on a personal vehicle. This work pioneers a reinforcement learning framework that detects novel observations and cautiously avoids them by reasoning about the neural network's predictive confidence.
Click for presentation and ICRA '19 paper .
Instruction of robots through expert programming is expensive. This work creates a cheap instruction method by allowing non-technical users to control a robot in real-time through an afforable motion capture suit.
.