Actively learning experimental designs in terrestrial climate science

Introduction

While land-atmosphere exchanges of carbon, water, and energy are key to understanding changes in the Earth system, we still fundamentally lack a methodology to obtain representative estimates of these surface fluxes at the scale of a single grid cell of an Earth System Model (typically 10-100 km), let alone for a wider region.

Project Overview

ACTIVATE combines an observing system consisting of:

• A swarm of drones carrying meteorological sensors and gas analyzers
• Mobile and stationary flux towers
• Satellites

It fuses their observations with different land-atmosphere models using data assimilation methods.

Methodology

ACTIVATE will develop an adaptive Bayesian Experimental Design framework to:

1. Generate maximally informative observation strategies for expensive data collection.
2. Adaptively reposition drone swarms during a flight as new observations become available to optimally infer surface fluxes in the landscape.

Demonstration

We will demonstrate the framework in three scenarios:

1. Idealized synthetic experiments.
2. Managed and industrial sites with known flux hotspots.
3. Targeted high-resolution simulations in poorly represented regions with expensive models that explicitly resolve subgrid-scale processes in Earth System models.

Application

We will apply the ACTIVATE framework around existing observatories in vulnerable arctic regions, where the lack of strong observational constraints from state-of-the-art observing systems is particularly apparent and problematic.

Expected Outcomes

ACTIVATE will produce:

• Unprecedented observational datasets for new model developments in some of the most data-sparse regions on Earth.
• Uncertainty-aware parameter estimates for critically unconstrained processes.
• A pioneering active experimental design framework for terrestrial observing systems.

Broader Vision

The broader vision of ACTIVATE is to develop active learning capabilities for improved data assimilation in models to elevate our understanding of land-atmosphere interactions across spatio-temporal scales.