Integrated assessment models (IAMs) are the dominant tools for projecting mitigation scenarios. However, IAM-based scenarios often face challenges such as modelling biases and large computational burden. Here we develop a deep learning framework to generate key variables through synthetic mitigation scenarios aligned with the Sixth Assessment Report (AR6) Scenarios Database. By analysing 1,202 scenarios from a diverse set of IAMs, we select key drivers that enable a more detailed sectoral representation. Next, we trained three generative deep learning models to produce 30,000 synthetic scenarios at low computational cost across various IPCC AR6 climate categories, replicating variable distributions and correlations while also demonstrating physical consistency in power sector variables through internal validation checks. We found that the variational autoencoder achieved the highest label transferring accuracy among three frameworks. This study illustrates the potential of deep learning to complement IAM approaches and provides a basis for handling complex mitigation scenario generation tasks.
