Reservoir Computing as a Tool for Climate Predictability Studies

Balasubramanya Nadiga2021 · DOI

TL;DR

This paper introduces Reservoir Computing (RC), a machine learning technique, as a powerful alternative to the traditional Linear Inverse Modeling (LIM) for climate predictability studies. RC demonstrates improved predictive skill, especially in scenarios with limited data, by effectively capturing nonlinear dynamics in climate systems.

What Problem Does This Paper Address?

Understanding and predicting climate variability, particularly "predictability of the first kind" (internal climate system variability), is a fundamental challenge in climate science. While comprehensive climate models are crucial, their computational cost limits large-scale ensemble predictions. Reduced-order models, like the Linear Inverse Model (LIM), offer a more computationally feasible approach by capturing essential interactions. However, LIM, being linear, may struggle to fully represent the inherently nonlinear and chaotic nature of climate dynamics, potentially limiting its predictive accuracy. This paper investigates whether a nonlinear approach, specifically Reservoir Computing (RC), can overcome these limitations and improve upon LIM's predictive skill.

Methods

The study employs Reservoir Computing (RC), a type of Recurrent Neural Network (RNN) where the internal "reservoir" weights are randomly initialized and fixed. Training then focuses solely on a linear "readout" layer. This approach leverages the reservoir's ability to create complex, nonlinear transformations of the input history. The authors compare RC's performance against the established Linear Inverse Model (LIM) approach. LIM involves inferring a linear dynamical system from data, typically in a reduced dimensional space (e.g., using Empirical Orthogonal Functions - EOFs). The study also explores the impact of data availability (long vs. short training periods) and the number of retained EOFs on the predictive skill of both methods.

Datasets

The primary dataset used is the pre-industrial control simulation (piControl) of the Community Earth System Model version 2 (CESM2). Specifically, the sea surface temperature (SST) data from the North Atlantic region over an 800-year period is utilized. This dataset is publicly available through the CMIP6 archive.

Key Insights and Ideas

  • Nonlinearity Matters: The inherent chaotic and nonlinear nature of climate dynamics means that linear models like LIM, while useful, may not capture the full picture of predictability. RC's ability to inherently handle nonlinearities offers a significant advantage.
  • Limited Data Advantage: RC shows a marked improvement in predictive skill over LIM when training data is scarce. This suggests RC is more robust to data limitations, a crucial factor in many real-world climate prediction scenarios.
  • Robustness of RC: The RC approach demonstrates consistent performance across different settings, including varying amounts of training data and numbers of EOFs, highlighting its general applicability.
  • Potential for Emulation: While not the primary focus, the paper hints at RC's potential for developing climate emulators – models that can continue system evolution on the attractor long after reference trajectories diverge, as demonstrated in the Lorenz-63 system.

Key Results

  • Improved Skill: In scenarios with limited training data, RC consistently outperforms LIM in terms of both Normalized Root Mean Square Error (NRMSE) and Anomaly Correlation Coefficient (ACC). For instance, at a 1-year lead time with limited data, RC shows significantly smaller errors and higher correlation than LIM.
  • Comparable Skill with Plentiful Data: When training data is plentiful (e.g., 480 years), RC's skill is comparable to, and marginally better than, LIM. This suggests that while LIM is effective in data-rich environments, RC offers a more robust and often superior alternative.
  • EOF Influence: The number of retained EOFs influences the performance of both methods. RC generally performs well across a range of EOFs, while LIM's performance can be more sensitive, especially in limited data regimes.
  • Non-monotonic Behavior in RC: Interestingly, RC exhibits a non-monotonic behavior in skill as training data is limited, with a dip in accuracy at intermediate training lengths before improving again. This suggests a transition between strongly nonlinear and more linear regimes.

Limitations and Future Work

  • Parameter Optimization: The study acknowledges that the RC parameters were not exhaustively optimized, and a more thorough search could potentially yield even better results.
  • Non-monotonic Behavior Explanation: The observed non-monotonic behavior of RC with limited data requires further investigation to fully understand its underlying causes.
  • Climate Emulator Development: The paper suggests that further development of the RC approach could enable its use as a climate emulator, a significant future research direction.
  • Generalizability: While the results are promising, the authors note that the improved skill of RC needs to be verified in other climate settings and for different predictability problems.
  • Reservoir Design: The optimal design of the reservoir itself remains an open question in RC research.

Summary updated

Abstract

<p>  Reduced-order dynamical models play a central role in developing our<br>  understanding of predictability of climate irrespective of whether<br>  we are dealing with the actual climate system or surrogate climate<br>  models. In this context, the Linear Inverse Modeling (LIM) approach,<br>  by helping capture a few essential interactions between dynamical<br>  components of the full system, has proven valuable in being able to<br>  provide insights into the dynamical behavior of the full system.</p><p>  We demonstrate that Reservoir Computing (RC), a form of machine<br>  learning suited for learning in the context of chaotic dynamics,<br>  provides an alternative nonlinear approach that improves on the LIM<br>  approach. We do this in the example setting of predicting sea<br>  surface temperature in the North Atlantic in the pre-industrial<br>  control simulation of a popular earth system model, the Community<br>  Earth System Model version 2 (CESM2) so that we can compare the<br>  performance of the new RC based approach with the traditional LIM<br>  approach both when learning data is plentiful and when such data is<br>  more limited. The useful predictive skill of the RC approach over a<br>  wider range of conditions---larger number of retained EOF<br>  coefficients, extending well into the limited data regime,<br>  etc.---suggests that this machine learning approach may have a use<br>  in climate predictability studies. While the possibility of<br>  developing a climate emulator---the ability to continue the<br>  evolution of the system on the attractor long after failing to be<br>  able to track the reference trajectory---is demonstrated in context<br>  of the Lorenz-63 system, it is suggested that further development of<br>  the RC approach may permit such uses of the new approach in settings<br>  of relevance to realistic predictability studies.</p>
BibTeX
@misc{nadiga2021reservoir,
  author = {Balasubramanya Nadiga},
  title = {{Reservoir Computing as a Tool for Climate Predictability Studies}},
  year = {2021},
  doi = {10.5194/egusphere-egu21-1561},
}
Figure