Generative Adaptation of Dynamics to Environmental Shifts via Weight-space Diffusion

Data-driven dynamics prediction often fails under environmental shifts, while traditional fine-tuning remains computationally prohibitive for hardware-constrained or data-scarce applications. We propose DynaDiff, a generative meta-learning framework that transitions the paradigm from gradient-based tuning or modulation to direct weight-space generation. Specifically, we first abstract expert weights as novel weight graphs, utilizing multi-head attention to explicitly capture topological coupling within weights. Subsequently, we design a functional loss to ensure that the generated models achieve consistency with expert models in physical behavior. Finally, we develop a dynamics-informed prompter that extracts cross-domain physical and spectral features from observation sequences to condition the diffusion model. Experiments demonstrate that DynaDiff boosts average prediction accuracy by 10.78\% over competitive baselines. Furthermore, by pre-constructing a model zoo of expert predictors, we amortize the fine-tuning overhead into a one-time offline cost, significantly boosting deployment efficiency in new environments.

AI models are increasingly used to predict how physical systems evolve over time, such as airflow, fluids, and weather. However, these models often work well only under the conditions they were trained on. When the environment changes, for example when fluid properties or external forces differ, the model may become inaccurate. A common solution is to retrain or fine-tune the model with new data, but this can be slow, expensive, and impractical when only a small amount of new data is available. This paper introduces DynaDiff, a method that adapts to new environments without retraining. Instead of repeatedly adjusting a model through trial and error, DynaDiff learns how good specialist models differ across many known environments. Given only a short observation from a new environment, it directly creates a new specialist model suited to that situation. To make this possible, DynaDiff represents the internal settings of each model in a structured way, learns which parts matter for the model’s behavior, and uses features from the observed dynamics to guide generation. Experiments on simulated physical systems and real-world wind data show that DynaDiff produces small, environment-specific models that are both accurate and efficient. It improves average prediction accuracy over strong competing methods while greatly reducing the cost of adapting to new environments. This makes it promising for scientific and engineering applications where conditions change, data is limited, or computing resources are constrained.