A Multitask Learning View on the Earth System Model Ensemble

Andre R. Goncalves, Fernando J. Von Zuben, Arindam Banerjee

Research output: Contribution to journalArticlepeer-review

Abstract

Earth system models (ESMs) are based on physical principles that are intended to emulate climate behavior. They're the primary mechanisms for obtaining projections of future conditions under different climate change scenarios. Because ESMs rely on the distinct modeling of certain physical processes and initial conditions, different ESMs can produce different responses for the same external forcing. Researchers consider climate projections based on ensembles of climate models with the goal of getting better accuracy and reduced uncertainty. The authors look at the problem of combining ESMs from a multitask learning (MTL) perspective, where ESM ensembles for all regions are performed jointly. By taking advantage of commonalities among regions, an MTL approach is expected to improve prediction in individual regions. The authors consider the problem of constructing ensembles of regional climate models for land surface temperature projections in South America. Their MTL algorithm produced more accurate predictions than existing methods for the problem.

Original languageEnglish (US)
Article number7274259
Pages (from-to)35-42
Number of pages8
JournalComputing in Science and Engineering
Volume17
Issue number6
DOIs
StatePublished - Nov 1 2015
Externally publishedYes

Keywords

  • Earth system model
  • multimodel ensemble
  • multitask learning
  • Scientific computing
  • structure learning

ASJC Scopus subject areas

  • Computer Science(all)
  • Engineering(all)

Fingerprint Dive into the research topics of 'A Multitask Learning View on the Earth System Model Ensemble'. Together they form a unique fingerprint.

Cite this