Trading off Controllability and Realism in Generative Physics Emulators

Machine-learned emulators trained on numerical models' outputs inherit systematic biases relative to real-world observations. While adapting these emulators toward observations can enhance their realism, it often reduces their diversity from a carefully constructed ensemble of simulations. This creates a fundamental \textit{controllability-realism tradeoff}, where gaining accuracy comes at the cost of losing the distinct identities of the underlying simulators. In this work, we formalize this tradeoff for conditional generative emulators and introduce \textit{condition-to-condition guidance}. Our method is an inference-time approach that interpolates between simulator and observational conditional score functions controlled by a guidance scale $\gamma$. We further provide a principled procedure for selecting an optimal $\gamma ^\star$. We evaluate our framework on a climate science use case and compare it with a classical bias correction method.