GP Atmospheric Inverse Modeling

Papers

1. Method Paper

Applying Gaussian Process Machine Learning and Modern Probabilistic Programming to Satellite Data to Infer CO2 Emissions

Paper link: https://pubs.acs.org/doi/10.1021/acs.est.4c09395.

Sponsor: NASA Carbon Cycle Science Program (80HQTR21T0101)

The code (Jupyter Notebook) implements a GPyTorch-based Gaussian Process (GP) marginal log-likelihood methodology.  

The data and code are available to the public; if the data are used in a publication, please let us know at sjeong@lbl.gov. 

Code and Input Data

Code

GP GPyTorch implementation (Jupyter Notebook)

Input Data

• y (y.csv): The target variable [ppm] representing the (observed) data points that the model aims to predict.

• K (K.csv): Matrix (note: this is not kernel but 𝐾𝑥 in the paper text) utilized for the mean model, based on perturbed GC-model predictions [ppm] categorized by sector. These data points are pre-multiplied by the parameter 𝜆, which scales the influence of each sector on the total predictions. Thus, (𝐾𝑥)𝜆 constitutes the mean function in the GP model.

• X (X.csv): Matrix of input features consisting of geographical coordinates (longitude and latitude, all in normalized UTM 10N) and temporal data (time), which are used to structure the spatial and temporal dimensions of the model.

2. Application Paper

State-wide California 2020 Carbon Dioxide Budget Estimated with OCO-2 and OCO-3 satellite data, Atmospheric Physics and Chemistry

Paper link: https://acp.copernicus.org/articles/25/8475/2025/

Sponsor: NASA Carbon Cycle Science Program (80HQTR21T0101)