Towards More Accurate and Adaptable Surrogate Models

Abstract

Physics-based models are essential for understanding and predicting environmental change, but their high computational cost limits their use in rapid forecasting and scenario analysis. This study explores data-driven surrogate models for approximating physics-based models to reduce computational demands without sacrificing accuracy. Building on an existing surrogate modeling framework, we aim to: (1) enhance model performance in a coastal case study through alternative dimensionality reduction and regression techniques, (2) improve realism by integrating observational data into latent space representations, and (3) evaluate the framework’s transferability through application to an urban case study. The results from the coastal case indicate that surrogate models using autoencoders and LSTM networks improve predictive performance by approximately 20% compared to the baseline implementation of PCA and linear regression, as measured by RMSE. This improvement is primarily attributed to the LSTM model. However, the autoencoder also demonstrated potential, such as its ability to incorporate observational data into the latent space. The urban case demonstrated the framework’s potential, but revealed new challenges. While predictions were acceptably accurate in most instances, model instability suggested that the dataset was insufficient and that alternative methods may be more suitable. Overall, the results highlight the strong potential of surrogate modeling as a computationally efficient complement to physics-based models. This capability enables rapid evaluation of climate scenarios and real-time forecasting. Future work should focus on incorporating extreme events into training objectives and tailoring model complexity to specific application needs.