See my previous research page for more details about my past research.
Physics-Guided Machine Learning for Climate Modeling
Understanding and combating the climate crisis is of paramount importance. The challenge of simultaneously simulating clouds and planetary-scale winds has been a key reason for uncertainty in future climate predictions, and it is unlikely that we will routinely run such simulations before 2050. Machine-learning algorithms trained on storm-resolving models run for shorter time periods can mimic the statistical effects of fine-scale clouds in less expensive climate models, and this is already accelerating climate modeling. However, statistical algorithms on their own have deficiencies that prevent their adoption by the climate community. Physical (theory-driven) modeling and machine-learning (data-driven) modeling have traditionally been treated as independent approaches, but recent progress in stochastic optimization allows training of machine-learning models with both flexible architectures and loss functions that can incorporate physical knowledge.
Our research has combined atmospheric physics and deep learning to address key shortcomings of neural-network (NN) models of convection and clouds because NNs are powerful non-linear regression tools, well-suited for high-dimensional problems and large datasets. First, unconstrained NNs typically violate mass and energy conservation laws, leaking energy at a rate that compromises long-term climate predictions. We developed methods to modify a NN’s architecture so as to enforce mass and energy conservation to within machine precision, which we then generalized to enforce non-linear constraints in any NN emulating a physical system. Second, NNs are hard to interpret and often cause instability when coupled to atmospheric fluid dynamics. We customized machine learning interpretability tools to improve the transparency and stability of NN models of convection. Third, NNs make large errors when evaluated out-of-distribution, e.g. in the Tropics of a warmer climate. We demonstrated that physically rescaling a NN’s inputs and outputs allowed the NN to make reasonable predictions in a different climate and in regions of the atmosphere it had not been trained on. Most recently, we formalized this framework to help machine learning algorithms maintain high accuracy across a wide range of climates and geographies in three distinct climate models.
AI for Tropical Meteorology
More coming soon, see our AI for Good webinar in the meantime.
Data-Driven Scientific Discovery
More coming soon, see our work on improved understanding through equation discovery and generative modeling in the meantime.
(2014-2020) Interaction between Water Vapor, Radiation and Convection in the Tropics
The interaction between the fast ascending motion of light air that produces clouds and storms (atmospheric convection) and large-scale winds is a primary source of uncertainty in numerical simulations of the atmosphere, impeding our understanding of the climate. At the beginning of my Ph.D., a salient problem caught my attention: When high-resolution atmospheric models were run to radiative-convective equilibrium (the simplest model of the tropical atmosphere capable of realistic temperature and water vapor predictions), strong storms spontaneously clustered together to form compact moist regions. While the research community agreed this “self-aggregation” of convection was not a numerical artifact, its physical mechanisms and applicability to the real world were still disputed.
Motivated by this opportunity to close the gap between numerical simulations, theory, and observations, we used simple radiative models to hypothesize that the observed atmosphere was humid enough for “self-aggregation” to occur almost everywhere in the real Tropics. The inability of simple theories to capture the spatial variability of tropical atmospheric water sparked my interest in data science; instead of proposing an additional theory for the size of moist and dry regions in the Tropics, we adapted spectral methods to quantify the role of radiation, surface fluxes, and advection in organizing water vapor objectively at each spatial scale from terabytes of cloud-resolving model, global storm-resolving model, and observational data. Systematically bridging atmospheric physics theory and observations via the creation of diagnostic data analysis tools is now a cornerstone of my research philosophy, and we recently leveraged theory from statistical physics to formulate a new index that can quantify the convective aggregation of both observed tropical rainbands and idealized storm clusters.
Atmospheric Water Dynamics
Climate Impacts & Risk Analysis
More coming soon, see our submission to the 2018 ClimateChanged@MIT competition in the meantime.
Perspectives and Literature Reviews
More coming soon, see our book chapter on Machine Learning for Clouds and Climate in the meantime.