Fire is a crucial factor in terrestrial ecosystems playing a role in disturbance for vegetation dynamics. Process-based fire models quantify fire disturbance effects in stand-alone dynamic global vegetation models (DGVMs) and their advances have incorporated both descriptions of natural processes and anthropogenic drivers. Nevertheless, these models show limited skill in modeling fire events at the global scale, due to stochastic characteristics of fire occurrence and behavior as well as the limits in empirical parameterizations in process-based models. As an alternative, machine learning has shown the capability of providing robust diagnostics of fire regimes. Here, we develop a deep-learning-based fire model (DL-fire) to estimate daily burnt area fraction at the global scale and couple it within JSBACH4, the land surface model used in the ICON ESM. The stand-alone DL-fire model forced with meteorological, terrestrial and socio-economic variables is able to simulate global total burnt area, showing 0.8 of monthly correlation (rm) with GFED4 during the evaluation period (2011-15). The performance remains similar with the hybrid modeling approach JSB4-DL-fire (rm=0.79) outperforming the currently used uncalibrated standard fire model in JSBACH4 (rm=-0.07). We further quantify the importance of each predictor by applying layer-wise relevance propagation (LRP). Overall, land properties, such as fuel amount and water content in soil layers, stand out as the major factors determining burnt fraction in DL-fire, paralleled by meteorological conditions over tropical and high latitude regions. Our study demonstrates the potential of hybrid modeling in advancing fire prediction in ESMs by integrating deep learning approaches in physics-based dynamical models.