Earth System Models provide important insight into global climate dynamics. These models often require large computational resources to run, inhibiting accessibility and exploration of a wide range of climate-related scenarios. Machine learning can help by creating an emulation of an aspect of an ESM to enable less expensive scenario simulation. I use a Long Short-Term Memory model to emulate forest carbon dynamics in the Community Earth System Model 2 in order to understand the impact of wood harvest on carbon stocks in the Canadian Boreal forest. To validate the emulation, I use available external datasets that explicitly quantify carbon stocks in soil and above-ground biomass. The emulation can predict CESM2 several carbon stock variables accurately (0.89 R$^2$ Score) and can be explained with important climatic relationships. I then create land-cover scenarios to simulate no wood harvest for the years 1984-2019. These scenarios show that 584 Mt C were lost to wood harvest over this period, with an additional 172 Mt C attributed to regrowth from wood harvest over the same period. The LSTM model I use in this study provides a more flexible approach to investigating land-use change impacts on carbon stocks by harnessing the power of both machine learning models and process-based ESMs. This approach can help understand land-use change scenarios that are not considered in large inter-model comparison efforts.