Technical note: Deep learning for creating surrogate models of precipitation in Earth system models

Weber, Theodore R.; Corotan, Austin; Hutchinson, Brian; Kravitz, Ben; Link, Robert

doi:10.5194/acp-20-2303-2020

Cited by 17 publications

(12 citation statements)

References 41 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…One approach is to emulate the dynamical evolution of these numerical models directly, and this has been explored for both weather [17,18] and climate [19–21]. While these are obviously very early efforts in this direction, they demonstrate that developing machine learning models that can compete with their traditional counterparts in numerical weather prediction is extremely challenging, and extending this to climate time scales even more so, especially given the difficulty in maintaining the energy and mass conservation required for a stable simulation.…”

Section: Climate Emulationmentioning

confidence: 99%

Machine learning for weather and climate are worlds apart

Watson-Parris

2021

Phil. Trans. R. Soc. A.

View full text Add to dashboard Cite

Modern weather and climate models share a common heritage and often even components; however, they are used in different ways to answer fundamentally different questions. As such, attempts to emulate them using machine learning should reflect this. While the use of machine learning to emulate weather forecast models is a relatively new endeavour, there is a rich history of climate model emulation. This is primarily because while weather modelling is an initial condition problem, which intimately depends on the current state of the atmosphere, climate modelling is predominantly a boundary condition problem. To emulate the response of the climate to different drivers therefore, representation of the full dynamical evolution of the atmosphere is neither necessary, or in many cases, desirable. Climate scientists are typically interested in different questions also. Indeed emulating the steady-state climate response has been possible for many years and provides significant speed increases that allow solving inverse problems for e.g. parameter estimation. Nevertheless, the large datasets, non-linear relationships and limited training data make climate a domain which is rich in interesting machine learning challenges. Here, I seek to set out the current state of climate model emulation and demonstrate how, despite some challenges, recent advances in machine learning provide new opportunities for creating useful statistical models of the climate. This article is part of the theme issue ‘Machine learning for weather and climate modelling’.

show abstract

Section: Climate Emulationmentioning

confidence: 99%

Machine learning for weather and climate are worlds apart

Watson-Parris

2021

Phil. Trans. R. Soc. A.

View full text Add to dashboard Cite

show abstract

“…For example, machine learning techniques (e.g., convolutional neural networks, deep learning, etc.) that have been applied to detect patterns in climate models (e.g., Chen et al, 2020; Weber et al, 2020; Yu et al, 2020) could greatly aid in the estimation of precipitation at ungauged locations. Accordingly, the use of high‐performance computing (HPC) is essential for extensive, expected amounts of data.…”

Section: Conclusion: Advances Gaps and The Challenges Of Grid Creationmentioning

confidence: 99%

From rain to data: A review of the creation of monthly and daily station‐based gridded precipitation datasets

Serrano‐Notivoli

Tejedor

2021

WIREs Water

View full text Add to dashboard Cite

Monthly and daily gridded precipitation datasets are one of the most demanded products in climatology and hydrology. These datasets describe the high spatial and temporal variability of precipitation as a continuous surface and for defined periods. However, due to the complex characteristics of precipitation, it is difficult to obtain accurate estimations. Thus, the creation of a gridded dataset from observations requires the comprehensive and precise application of quality control, reconstruction, and gridding procedures. Yet, despite multiple advances, most of the gridded datasets created and published since the mid-1990s to the present use a wide variety of techniques, methods, and outputs, which can completely change the final representativity of the data. It is, therefore, critical to provide general guidelines for the development of future and more robust gridded datasets based on the data characteristics, geographical factors, and advanced statistical techniques. We identified gaps and challenges for near-future perspectives and provide guidelines for implementing improved approaches based on the performance of 48 products. Finally, we concluded that, despite better spatial and temporal resolutions, data access, and data processing capabilities, observational coverage remains a challenge. Moreover, scientists should adopt tailored strategies to improve the representativity and uncertainty of the estimates.

show abstract

“…Surrogate models can readily incorporate rich datasets, such as from SAIL [ Chen et al , 2020;Liu et al , 2020;Mital et al , 2020], and can quickly discern obvious and non-obvious relationships that impact precipitation and radiation without having to first sleuth out WRF model errors. Surrogate models, such as deep neural networks, have already been shown to discern these relationships [ Weber et al , 2020]. While there are real risks of surrogate model over-training and challenges with model error diagnostics, surrogate models can quickly transfer learning from SAIL to other under-studied watersheds [ Chandrasekar, 2020].…”

Section: Narrativementioning

confidence: 99%

Reliable modeling and prediction of precipitation & radiation for mountainous hydrology

Feldman

2021

View full text Add to dashboard Cite

2018a,b,c], all of these models exhibit limited predictability in the date of peak snowpack timing and spring snowmelt rate across the Western United States. However, there is a large chasm between identifying these common problems and fixing the models so that, at the very least, they produce unbiased estimates of mountainous hydrology.Advances in mountainous hydrological science are needed to understand and improve the prediction of these systems from synoptic to decadal time-scales and with spatial scales ranging from meters to the entire watershed (potentially thousands of km) [ Bales et al. , 2006]. Such advances require coordinated modeling and observational data collection [ Viviroli et al., 2011]. Spatially-and temporally-resolved precipitation and radiation fields are extremely heterogeneous in complex high-altitude terrain (CHAT), and networks of point observations under-resolve the processes that contribute to this heterogeneity. Significant resources are being devoted by the DOE to resolve these processes in targeted watersheds, such as the East River Watershed in the

show abstract

Technical note: Deep learning for creating surrogate models of precipitation in Earth system models

Cited by 17 publications

References 41 publications

Machine learning for weather and climate are worlds apart

Machine learning for weather and climate are worlds apart

From rain to data: A review of the creation of monthly and daily station‐based gridded precipitation datasets

Reliable modeling and prediction of precipitation & radiation for mountainous hydrology

Contact Info

Product

Resources

About