A Real-Time, Auto-Regression Method for In-Situ Feature Extraction in Hydrodynamics Simulations

TL;DR

The paper tackles the data-generation bottleneck in hydrodynamics by introducing a lightweight real-time in-situ feature extraction method based on linear auto-regression trained with mini-batches during simulation. It jointly collects temporal and spatial data, fits curves, and tracks features like local extrema and inflection points to extract meaningful dynamics, all via a flexible API library. Empirical evaluations on LULESH material deformation and Castro WD mergers demonstrate high accuracy (approximately 94%–100%) and low overhead (roughly 0.1%–5%), with substantial potential for early termination that speeds up simulations by up to ~67%. The approach reduces data movement, enables timely feature discovery, and provides a practical framework for integrating in-situ analytics into large-scale hydrodynamic workflows.

Abstract

Hydrodynamics simulations are powerful tools for studying fluid behavior under physical forces, enabling extraction of features that reveal key flow characteristics. Traditional post-analysis methods offer high accuracy but incur significant computational and I/O costs. In contrast, in-situ methods reduce data movement by analyzing data during the simulation, yet often compromise either accuracy or performance. We propose a lightweight auto-regression algorithm for real-time in-situ feature extraction. It applies curve-fitting to temporal and spatial data, reducing data volume and minimizing simulation overhead. The model is trained incrementally using mini-batches, ensuring responsiveness and low computational cost. To facilitate adoption, we provide a flexible library with simple APIs for easy integration into existing workflows. We evaluate the method on simulations of material deformation and white dwarf (WD) mergers, extracting features such as shock propagation and delay-time distribution. Results show high accuracy (94.44%-99.60%) and low performance impact (0.11%-4.95%) demonstrating the method's effectiveness for accurate and efficient in-situ analysis.

Figures (8)

• Figure 1: Illustration of variable tracking and peak value detection. The gradients such as $k_1$, $k_2$, and $k_3$ are computed, and in this case the negative $k_3$ indicates the peak value is detected.
• Figure 2: Code to show the usage of APIs from our library framework to program feature extraction. The main computation is shown in lines 25 and 26. Bold functions are provided by the library framework.
• Figure 3: Illustration of blast wave propagation within a 3D domain. The wave originates at point O, with velocities at any position (x, y, z) computed based on their components along each axis. The arc surface represents the wavefront, where velocities (v1, v2, ..., v5) are equal due to spherical symmetry. This illustration is adapted from hornung2011hydrodynamics.
• Figure 4: Curve-fitting results on velocities at location 10 using lag values of 50 and 100 over 40%, 60%, and 80% of total iterations with a domain size of 30. Only error rates between 0.00 and 0.25 are shown in the figures.
• Figure 5: Distribution of velocity over timesteps 1 to 932 at locations 1 to 10. The velocity curves are aligned into the same figure by number of iterations. The threshold is used to determine the breaking-point of certain material as the feature to be extracted.