In-Situ Auto-Regressive Surrogate Modeling for Feature Extraction Using Ascent

Surrogate modeling is a crucial technique for accelerating scientific simulations by approximating complex computational processes with lightweight predictive models. This paper explores the use of Ascent, an in-situ visualization and analysis framework, to train an auto-regressive surrogate model for selected dynamic features during simulation runtime, enhancing fast and comprehensive feature extraction from scientific simulations. By leveraging Ascent’s in-situ capabilities, our approach continuously extracts temporal data from the simulation at runtime and incrementally updates the model’s weights by stochastic gradient descent (SGD) using both new data and its previous predictions. This enables real-time training of an auto-regressive model, ensuring accurate surrogate modeling while minimizing computational overhead and avoiding disruptions to the simulation workflow. We apply this method to benchmark simulations, including LULESH, Laghos, and Kripke, demonstrating its effectiveness in capturing key dynamic features while significantly reducing computational cost of simulation. Our results show that the trained surrogate model achieves an accuracy ranging from 97.74% to 99.56% in predicting simulation states, while accelerating the simulation ranging from 1.8× to 4.55×. This work highlights the potential of integrating in-situ machine learning with scientific computing, paving the way for more efficient and scalable simulation workflows.