Workshop insights: Machine learning for earth systen modeling and analytics

Helmholtz-Zentrum Hereon, Climate Service Center Germany (GERICS) and German Climate Computing Centre (DKRZ) co-organized a virtual workshop on “Machine Learning for Earth System Modeling and Analytics” with the support of Helmholtz AI. The workshop took place on 3 and 4 May 2021 with about 200 participants, of which about one third were from Helmholtz Centers and another third were international participants. The workshop called for contributions on applying machine learning (ML) methods particularly beyond computational modelling. This theme was well received and both presenters and audience shared interests in Earth observation, impact assessment and natural hazards.

Keynotes were given by Marc van den Homberg (510/Netherlands Red Cross) on leveraging AI for disaster risk management and Kelly Caylor (University of California, Santa Barbara (UCSB)) on ML with observational data and the need for labelled datasets. Several research activities that received support from Helmholtz AI, including voucher support, also presented, covering topics ranging from innovative remote sensing methods to near-surface temperature and air quality forecasting.

The plenary discussions shed light on progress on topics at the heart of the Earth system sciences community seeking to make progress beyond known application fields of ML. Discussion topics covered perspectives on the capabilities of ML to substitute or complement established approaches based on computational models and physical laws, which concluded that hybrid approaches informed by physical constraints may be the most promising long-term approach. This perspective also reflects the observation that data-driven approaches have already proven their potential and are moving beyond first demonstrators, but their benefit mostly lies in reducing computational costs. ML-based techniques that enhance capabilities to understand underlying physics in new ways are a promising focus for intensified research efforts.

Following the second keynote, the workshop concluded with a call for assembling high-quality labelled training datasets that are useful beyond single use cases. It would be much more economical and beneficial for wider uptake of ML methods if labelled datasets are created and curated that fit many of the subdisciplinary challenges and would, thus, help to overcome the bottlenecks currently perceived particularly for ML projects working on observational phenomena.

More information

Further information about the workshop including some of the presentations given are available on the workshop web page: https://www.dkrz.de/p/ml-workshop-2021