Enhancing CMIP6 climate predictions through machine learning

Abstract

This study examines how machine learning (ML) approaches can be applied to enhance the performance of CMIP6 multi-model ensembles (MME) for climate projections across ten vulnerable locations in India. The research evaluates traditional MME methods (simple mean) alongside ML models—Long Short-Term Memory (LSTM), Artificial Neural Networks (ANN), and Support Vector Regression (SVR) to predict precipitation (PCP), maximum and minimum temperature (TMAX and TMIN) under both scenarios SSP245 and SSP585. Key findings include performance improvement of ML models consistently outperforming traditional MME, with LSTM achieving the highest R2 values (e.g., 0.85 for precipitation in Location 3 under SSP245) and reduced RMSE and MAE. SVR and ANN also showed significant improvements, particularly in capturing extreme events and seasonal trends. Temperature Projections show that all methods performed well for temperature variables, with minor variations, as temperature trends exhibit less variability over time. Trend Analysis shows that the MME-mean revealed statistically significant increasing trends in all locations, while LSTM displayed high variability, and ANN provided more stable projections. SVR was less reliable for long-term trend detection. Entropy Analysis: Variability indices (SVIAE and SVIME) indicated that SVR and MME-mean exhibited higher variability, whereas LSTM and ANN produced more consistent results, especially at annual scales. The study concludes that ML-augmented ensembles, particularly LSTM, enhance the accuracy of climate projections, offering valuable insights for climate resilience planning in vulnerable regions. However, traditional MME remains robust for consensus-based trend analysis. These findings contribute to optimizing climate model ensembles for improved decision-making in adaptation strategies. Keywords: Climate Change, Extreme Events, Machine Learning, Climate Variability