Comparative evaluation of system identification methods for large-scale nonlinear climate models

Determine, through comprehensive benchmarking on dynamic climate models, which system identification and model-learning algorithms are most effective and scalable for large-scale, complex nonlinear dynamic networks, and develop standardized computational comparisons to assess their performance under high dimensionality, uncertainty, and partial observability.

Background

The tutorial frames climate models as high-dimensional nonlinear dynamical systems and highlights the relevance of control-theoretic tools, including system identification, for learning and calibrating such models. While the control literature offers a wide range of identification and learning techniques, their relative effectiveness at the scales and complexities of climate models remains unclear.

The authors point out a gap in the existing research: there is a lack of systematic, computationally grounded comparative studies of identification approaches on realistic, large-scale nonlinear networks. They propose that dynamic climate models—characterized by uncertainty, under-sensing, and strong nonlinear couplings—are ideal benchmarks for such evaluations. Addressing this gap would clarify which methods are best suited for the demands of climate-system applications and could guide future methodological development.

References

Relevant to this paper, we note that it is still virtually unclear what is the overwhelmingly better method to perform large-scale system identification in complex nonlinear dynamic networks. In short, there is also an absence of computational studies that focus on comparing various identification and model learning approaches.

— Climate Science and Control Engineering: Insights, Parallels, and Connections  (2504.21153 - Elsherif et al., 29 Apr 2025) in Section 7 (Tutorial Summary and Moving Forward), bullet “Comparative analysis of system identification algorithms for climate models”