Transforming Earth Science by Bridging Machine Learning & Physics
Abstract
With rapid advances in both computational power and access to relevant data, the practice of the Earth Sciences (indeed all sciences) is being dramatically transformed. Meanwhile, robust theories of watershed/hydrometeorological system behavior have largely proved elusive due to land-surface hydrology lying in the interesting intermediate realm of "organized complexity" that resists robust analytical or statistical treatment. Since as early as 1995, ANN type models have been shown to provide better streamflow forecasts than state-of-the-art watershed models. More recently several large-sample studies have disconcertingly shown that Machine Learning (ML) can dramatically outperform Conceptual/Physics-Based models of land surface water-and-energy dynamics when making predictions out-of-sample.
The challenge, therefore, is to somehow synthesize the structural inference (pattern learning) strengths of ML with the (mass, energy and momentum) regularization strengths of Physics-Based (PB) Hydro-Meteorological modeling (and also the latter's ability to support improved conceptual/theoretical understanding). While ML can help us to extract pattern information directly from large volumes of available data of numerous kinds, and to detect relationships that might not otherwise have been imagined, PB allows us to impose known physics-based constraints (symmetry, conservation, locality), represent "causality" (long believed to be necessary to make robust out-of-sample predictions), and estimate behaviours of system attributes that are typically not part of the available data set (such as system state variables). This talk will examine the possibility of developing a transformed approach to the Earth Sciences in which Hybrid ML-PB models are used to both (1) learn from data, and (2) support conceptual understanding. We will conclude with a discussion of some possible ways towards achieving the scientific "holy-grail" of developing modeling systems that can automatically learn from their interactions with data.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFM.H33A..01G
- Keywords:
-
- 1805 Computational hydrology;
- HYDROLOGY;
- 1875 Vadose zone;
- HYDROLOGY;
- 1879 Watershed;
- HYDROLOGY;
- 1942 Machine learning;
- INFORMATICS