Assessing the Economic and Environmental Consequences of Climate Change

Uncertainties in estimating Earth’s future climate stem from both inaccuracies in our models and the vast array of possible choices that society will make in the intervening years. One of the most pressing uncertainties in climate modeling is that of the effect of anthropogenic aerosol, particularly through their interactions with clouds. Here I will introduce a general earth system emulation framework which leverages advances in machine learning and describe its application to the emulation of entire climate models for the reduction of this uncertainty. I will also demonstrate how such emulation can be used to better approximate the climate response to different anthropogenic forcing agents in order to aid in their detection and attribution, and in the exploration of different future emissions pathways.

There is a  growing demand to quantify parametric uncertainty as well as economic and climate uncertainty on the climate policies to tackle global warming. To investigate parametric uncertainty and nonlinear interactions among the uncertain model parameters, as well as learning about the climate equilibrium sensitivity, we develop a high-dimensional stochastic climate-economy model that propagates parametric uncertainty as pseudo-states. We approximate all equilibrium functions using deep equilibrium nets, which we show to be up to $mathcal{O}(10^5)$ faster than other state-of-the-art methods. To limit the number of model evaluations to obtain convergent statistics,  we further interpolate the outcomes of the cheap-to-evaluate surrogate model employing  Gaussian process regression in combination with  Bayesian active learning,  from which we estimate the  Sobol’  indices and univariate effects. 

We study the implications of model uncertainty in a climate-economics setting with three types of capital: “dirty” capital that produces carbon emissions when used for production, “clean” capital that generates no emissions but is initially less productive than dirty, and knowledge capital that increases with R&D investment and can lead to technological innovation in green sector productivity. To solve our high-dimensional, non-linear model framework we implement a neural-network-based global solution method. We show that there are first-order impacts of model uncertainty on optimal decisions and social valuations in this integrated climate-economic-innovation framework.  Specifically, accounting for interconnected uncertainty over climate dynamics, economic damages from climate change, and the arrival of a green technological change leads to important adjustments to investment in green capital and R&D investment in anticipation of potential technological change and the revelation of climate damage severity.

NVIDIA is committed to helping the scientific community address climate change and the energy transition. In Nov. 2021, our CEO announced the Earth-2 initiative, which aims to build digital twins of the Earth. Two central goals of this initiative are: (i) Computational: enable high-resolution hybrid climate-ML predictions with credible cloud physics via NVIDIA Modulus; and (ii) Societal: nimbly serve interactive, high-fidelity, high-resolution climate predictions via NVIDIA Omniverse. Next-generation km-scale climate simulations are the most credible ground truth envisioned for future climate change, but such simulations are prohibitively expensive, account for sizable carbon production into the atmosphere, and generate unmanageable amounts of output given that hundreds of diverse climate trajectories are needed to sample the long tails of risk. Therefore, the above-mentioned goals depend on achieving orders-of-magnitude speedup and data compression via a combination of advances in AI and computing technologies. 

NVIDIA sees potential to accelerate this work through collaborations with the climate science community and by bringing to bear its deep learning and data engineering expertise and HPC system deployments. Professional ML engineering pipelines – like those behind NVIDIA’s Modulus and Omniverse – have not been widely exploited in hybrid physics-ML climate simulation and Earth system emulation. For weather, the potential is already clear: AI-driven weather emulators show remarkably skillful data-driven high-resolution global weather predictions. Eventually the goal is for similar, high-fidelity digital twins of future Earth system processes to allow climate scientists, and non-expert users to interact more easily with the wealth of information in km-scale climate simulations and ML model predictions, facilitating planning and decision-making in climate monitoring, modeling, mitigation, and adaptation. We conclude with a roadmap of Earth-2 and its climate goals, and the engineering innovations required for the breakthroughs that building digital twins of the Earth demands.