Wildfires in the western United States (US) have become increasingly frequent and widespread, causing significant harm to human society. In this study, we used an interpretable hybrid machine learning (ML) model to understand and predict occurrences of large fires, explicitly incorporating the controls of fuel flammability, fuel availability, and human suppression effects on fires. Our model achieved a notable F1-score of 0.846 ± 0.012, outperforming traditional process-driven fire danger indices (models) and four commonly used ML models by up to 40% and 9%, respectively. In addition, the relationships between fires and their drivers, identified by our hybrid model, were aligned closer with established fire physical principles. Our further data analysis highlighted the distinct behaviors of human- versus lightning-ignited large fires in the western US. We quantified the respective vapor pressure deficit thresholds (VPDt) and associated fire weather risks for each fire type. Human-ignited large fires exhibited consistently lower VPDt and a greater annual number of flammable days across all western US ecoregions, with a 21% increase in flammable days for human-caused fires from 1979 to 2020. This increase outpaced that of lightning-caused fires. Further analysis highlighted the dominant role of greenhouse gas (GHG) emissions in modulating VPD and fire, which accounted for 81% of the increase in human-related flammable days. Overall, our results emphasize the necessity of trustworthy AI algorithms to tackle the challenges posed by rising compound climate extremes and their linkage to large wildfires. We also highlight the distinctly different characteristics for modeling human- vs. lightning-ignited large fires.