The Global Change Analysis Model (GCAM) simulates complex interactions between coupled Earth and human systems. In doing so, it provides insights into the effects of policy choices and climate on the land, water, and energy sectors. Because of the complexity of these systems and their interactions, it can be costly to run GCAM simulations at scale, which is required for ensemble studies that explore uncertainty in the model parameters and outputs. The same complexity can make it difficult to analyze GCAM and its behavior. A differentiable emulator with similar predictive characteristics, but greater efficiency, could facilitate novel scenario discovery, uncertainty exploration of GCAM, and discovery of pathways (i.e., combination of input factors) leading to desired outcomes. We present here a deep-learning emulator of GCAM with these capabilities. As a first use case, we train the emulator on an existing large ensemble designed to study factors affecting the adoption of renewable energy technologies. We investigate the effects of different sampling strategies for generating training data, the effect of training set size on model performance, and on the ability to identify novel pathways, as we feed the emulator values of parameters and their combination previously untested through GCAM itself. We demonstrate the ability to predict over 22,000 GCAM output values on a held out test set with greater than 0.98 R^2, and find that the agreement on input-output sensitivity between the emulator and GCAM is greater than 0.8 R^2. The end-to-end differentiability of the emulator enables efficient search over the input space to identify pathways that maximize target outcomes; as a proof of concept, we are able to reliably identify pathways leading to high adoption rates of renewable energy by 2050.