Sandy Hardian Susanto Herho 

This study evaluates machine learning emulators for modeling long-term atmospheric CO₂ evolution by comparing Random Forests (RF) and Multilayer Perceptrons (MLP) in replicating cGENIE Earth System Model outputs over a one-million-year timescale. Using one-year pulse emission experiments spanning 1,000-20,000 PgC with outputs tracked for 10⁶ years, we assessed emulator performance across multiple carbon cycle timescales. The RF emulator achieved superior accuracy (mean R² = 0.998 ± 0.001) and computational efficiency, reducing simulation time from weeks to seconds, while MLP showed lower performance (mean R² = 0.890 ± 0.015). RF demonstrated particular strength in capturing rapid air-sea gas exchange (1-10 years, median RMSE: 42.3 ppmv), ocean mixing (10-100 years, median RMSE: 23.4 ppmv), carbonate compensation (100-1,000 years, median RMSE: 15.6 ppmv), and long-term weathering feedbacks (> 1,000 years, median RMSE: 18.9 ppmv). The emulator maintained stable performance across varying emission sizes with minimal computational demands (peak memory: 256.8 MB). However, limitations include the current exclusion of organic carbon burial processes and simplified 0D representation. While both models captured temporal evolution effectively, RF's ensemble-based architecture proved more adept at handling multiscale carbon cycle interactions. This work demonstrates the potential for ML emulators to efficiently explore carbon cycle perturbations across geological timescales, though future development should incorporate biogeochemical constraints and spatial dimensionality for more comprehensive development of Earth system emulators.