Advances in artificial intelligence for climate science methods applications and evaluation challenges

To synthesize recent advances in artificial intelligence (AI) (machine learning (ML)/deep learning (DL)) for climate science across climate modeling, extreme-weather analysis, renewable-energy optimization, emissions monitoring, and climate adaptation. Using a PRISMA-aligned protocol, we surveyed studies published between 2015 and 2025 across major scholarly databases. We extracted task definitions, datasets, model classes, evaluation metrics, baselines, validation design, uncertainty treatment, openness (code/data), and energy/CO $$_2$$ reporting. Evidence is organized by application domain and by methodological rigor, emphasizing transferability and operational relevance. Strong evidence exists for improved short- to medium-range forecasting, nowcasting, bias correction, and anomaly detection. Physics-guided and hybrid models, including PINNs and neural operators, increasingly rival operational numerical weather prediction (NWP) systems in targeted settings. However, robust external validation, out-of-distribution testing, uncertainty quantification, and transparent reporting of code and computational footprint remain inconsistent. AI enhances but does not replace physically based climate modeling. We contribute a field-level synthesis that combines PRISMA screening with a study-quality checklist to foreground evaluation rigor, reproducibility, and sustainability. Adoption of physics-guided architectures, probabilistic prediction, regime-aware validation, and carbon-aware benchmarking is essential for decision-grade climate AI.