From Physics-Informed Models to Deep Learning: Reproducible AI Frameworks for Climate Resilience and Policy Alignment

ICLR 2026 Conference Submission22767 Authors

Keywords: Generative AI, LLMs, Artificial Intelligence, Deep Learning, Machine Learning

TL;DR: This paper presents a reproducible AI framework for climate resilience and policy alignment.

Abstract:

This paper investigates the multifaceted role of artificial intelligence (AI) in advancing climate resilience, with a specific focus on its
alignment with the complex and often contradictory US federal policy landscape. Through a reproducible research framework, a problem is formulated to predict localized temperature anomalies using open-source data from Berkeley Earth. The methodology employs a comparative analysis of a simpler, physics-informed model, Linear Pattern Scaling (LPS), against a more computationally-intensive deep learning model. The findings demonstrate that while deep learning shows promise for complex variables, such as precipitation, simpler models may offer superior performance and significantly lower computational overhead for temperature prediction, a critical point supported by recent MIT research. The report then synthesizes these technical findings with a detailed analysis of US federal policy, revealing a core conflict: the clean energy incentives of the Inflation Reduction Act (IRA) are in direct tension with the deregulatory, pro-fossil fuel stance of the current administration’s AI Action Plan. The report concludes that the net impact of AI on climate change is not predetermined, but is contingent on the creation of enabling conditions—including policy alignment, a focus on computational efficiency, and proactive measures to mitigate algorithmic bias and ensure climate justice.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 22767