• Anmol Chaure
• Ashok Kumar Behera
• Sudip Bhattacharya

As per the IPCC 6th assessment report, Anthropogenic activities continue to cause increasing and prolonged impacts on the Earth’s system. There is a growing need for tools and data driven models to develop adaptation plans, vulnerability assessments and resilience strategies. In various future socioeconomic scenarios fully coupled GCMs, (known as Earth System models) are the most used numerical models for climate change projection generations. These models have been the benchmark of climate modelling since their inception. However due to numerous parameters and an extensive spectrum of spatial and temporal scales, these models entail substantial computational costs which are usually executed as part of international experiments like CMIP (Coupled Model Inter-comparison Project). To circumvent this, one of the obvious solutions is the use of surrogate models of the GCMs. Often termed as emulators, surrogate models are designed to mimic the functionality of a GCM but still being very light in terms of computational as compared to a GCM model. Statistical surrogates have been prevalent for climate modelling since many years. However due to availability of huge volumes of observational, simulation and reanalysis climate data in the recent years, increase in the computational power of modern CPUs/GPUs and growing popularity of data driven modelling techniques, machine learning and deep learning based surrogate models for climate are becoming more prominent. In contrast to statistical surrogates, deep learning surrogates can capture complex non linear relationship between data, and can adapt to changing data patterns enabling them to capture even the minute dynamics of climate systems. Although having an advantage on some aspects, deep learning surrogates lack interpretability, can extrapolate values beyond the training points (lack of calibration) and cannot capture model uncertainty to a satisfactory extent. To mitigate these issues, GPR emerges as a viable candidate, as it uses a probabilistic approach to model the uncertainty estimates. GPR based models can be further tuned for better accuracy through the use of composite kernels. With the advancements in machine learning algorithms and increasing number of domains across various engineering streams employing GPR models, It is expected that better composite kernel design techniques will be introduced enabling higher accuracy and resolution in emulation techniques. The present work is a comprehensive summary of the potentials and limitations of GPR, as a surrogate for modelling climatic processes.

Earth System Science, Deep learning, Gaussian Process Regression, Climate Modelling

"Gaussian Process Regression for Climate Modeling: Potentials, Limitations, and Advances in Emulation Techniques", International Journal of Emerging Technologies and Innovative Research (www.jetir.org), ISSN:2349-5162, Vol.10, Issue 5, page no.a277-a287, May-2023, Available :http://www.jetir.org/papers/JETIR2305040.pdf

"Gaussian Process Regression for Climate Modeling: Potentials, Limitations, and Advances in Emulation Techniques", International Journal of Emerging Technologies and Innovative Research (www.jetir.org | UGC and issn Approved), ISSN:2349-5162, Vol.10, Issue 5, page no. ppa277-a287, May-2023, Available at : http://www.jetir.org/papers/JETIR2305040.pdf