Physics-Informed Neural Networks for Quantifying Solar-forcing-driven Variability in Earth’s Surface Temperature: Evidence Across Spatial Scales
Research in Astronomy and Astrophysics · 02.02.2027 · 27 (2) · 025011
DOI: 10.1088/1674-4527/ae7cc5 ↗Kurz gefasst
Fachartikel in der Zeitschrift Research in Astronomy and Astrophysics, online erschienen am 16.11.2026. Verfasst von Saleh Baltyuor und 9 weiteren Autor:innen. Die Arbeit nennt 37 Literaturangaben.
Abstract (Originalsprache)
Quantifying the terrestrial response of solar variability across the 11 yr Schwabe cycle remains an open challenge in solar-terrestrial physics, due to the dominance of anthropogenic warming signals and the limitations of conventional statistical approaches. We present a Physics-Informed Neural Network (PINN) constrained by the zero-dimensional energy balance equation C d T / d t = αS ( t ) + γ CO 2 ( t ) − βT ( t ), trained on 1488 months of detrended global surface temperature anomalies (1900–2023). By jointly minimizing data misfit and the energy balance residual, the PINN learns forcing parameters directly from observations while remaining physically constrained. The model achieves R 2 = 0.844 ± 0.002 on held-out test data, compared to R 2 = 0.178 for conventional multiple linear regression (MLR) applied to the same detrended dataset, demonstrating that physics-constrained temporal modeling provides both the predictive skill and the parameter stability that instantaneous regression and unconstrained temporal models cannot simultaneously achieve. The learned solar sensitivity α = 0.088 ± 0.001 is physically consistent with published estimates of solar climate sensitivity. Spectral analysis of the model-isolated solar component independently recovers the Schwabe solar cycle at ∼10.2 yr without prior specification of the expected period. The solar-forcing share, expressed as α /( α + γ ) = 15.5% ± 0.3%, is stable across five independent training realizations and quantifies the solar contribution to the total modeled radiative forcing sensitivity in the detrended residual record. Applying the same framework across spatial scales reveals systematic degradation of predictive skill from global to urban station scale ( R 2 = 0.844 to R 2 ≈ 0), while the solar share remains physically consistent (14.7–18.5% ± 0.5%) across all converging analyses. A historical Riyadh-city station independently recovers a dominant spectral peak consistent with the mean Schwabe period (11.0 ± 0.5 yr) with R 2 = 0.577 ± 0.060, confirming the presence of the solar signal in pre-urban Arabian Peninsula desert temperature records. The scale-dependent degradation in model performance demonstrates that anthropogenic urban processes elevate the effective thermal noise floor beyond the detectable amplitude of the solar signal, with implications for solar–terrestrial studies relying on urban ground-station datasets. An ablation study confirms that the energy balance physics constraint—rather than the LSTM temporal memory alone—is the primary driver of parameter stability and interpretability. Removing this constraint destabilizes the solar attribution from 15.3% ± 0.6% to 62% ± 32% across training seeds.
Zitieren
APA
Baltyuor, S., Elmhamdi, A., Liu, Y., Liu, Q., Oloketuyi, J., Hu, Y., Al-Mostafa, Z. A., Alruhaili, A., Hu, X., & Yang, S. (2027). Physics-Informed Neural Networks for Quantifying Solar-forcing-driven Variability in Earth’s Surface Temperature: Evidence Across Spatial Scales. Research in Astronomy and Astrophysics, 27(2), 025011. https://doi.org/10.1088/1674-4527/ae7cc5
Vancouver
Baltyuor S, Elmhamdi A, Liu Y, Liu Q, Oloketuyi J, Hu Y, et al. Physics-Informed Neural Networks for Quantifying Solar-forcing-driven Variability in Earth’s Surface Temperature: Evidence Across Spatial Scales. Research in Astronomy and Astrophysics. 2027;27(2):025011. doi:10.1088/1674-4527/ae7cc5
BibTeX
@article{Baltyuor2027,
author = {Baltyuor, Saleh and Elmhamdi, Abouazza and Liu, Yu and Liu, Qiang and Oloketuyi, Jacob and Hu, Yang and Al-Mostafa, Zaki A. and Alruhaili, Aied and Hu, Xing and Yang, Shangbin},
title = {Physics-Informed Neural Networks for Quantifying Solar-forcing-driven Variability in Earth’s Surface Temperature: Evidence Across Spatial Scales},
journal = {Research in Astronomy and Astrophysics},
year = {2027},
volume = {27},
number = {2},
pages = {025011},
doi = {10.1088/1674-4527/ae7cc5}
}
EUVIMED