History-aware adaptive reduced-order models via incremental singular value decomposition

2026-05-27 · Source: Takara TLDR - Daily AI Papers · Field: Science & Research — Engineering & Applied Sciences, Mathematics & Computational Sciences · Depth: Expert, medium

Summary

A new projection-based adaptive Reduced-Order Model (ROM) framework, utilizing incremental Singular Value Decomposition (iSVD), addresses the accuracy decline of ROMs when online dynamics diverge from offline training data. This "history-aware" iSVD method updates the model's basis through occasional full-order operator evaluations, ensuring updates propagate to reduced operators. Evaluated on the one-dimensional viscous Burgers equation, the Sod shock tube, and a stiff one-dimensional ten-species rotating detonation engine (RDE), iSVD consistently outperformed alternative basis adaptation rules. Specifically, for the RDE problem, the iSVD adaptive ROM improved both predictive accuracy and computational efficiency compared to the current Direct adaptive ROM baseline. Cost analysis indicates that obtaining correction snapshots from the full-order model is the primary online expense, with the iSVD update itself being negligible. This positions iSVD as an effective mechanism for online learning of reduced subspaces, extending ROMs' predictive capabilities significantly beyond initial training windows.

Key takeaway

For research scientists developing reduced-order models for complex dynamical systems, integrating history-aware iSVD can significantly extend model predictive horizons. You should consider implementing iSVD for online basis adaptation, especially in scenarios where system dynamics evolve beyond initial training data. Focus optimization efforts on minimizing the cost of full-order model interactions for correction snapshots, as this is the primary computational bottleneck. This approach offers improved accuracy and efficiency over existing adaptive ROM baselines.

Key insights

History-aware iSVD adaptively updates reduced-order models, improving accuracy and efficiency in dynamic simulations.

Principles

• Online basis updates enhance ROM accuracy.
• History-aware updates outperform instantaneous ones.
• Dominant cost is full-order model interaction.

Method

The method involves projection-based adaptive ROMs using iSVD. It updates the basis online with correction snapshots from occasional full-order operator evaluations, propagating changes to reduced operators.

In practice

• Apply iSVD for adaptive ROMs in nonlinear problems.
• Use full-order model evaluations for correction snapshots.
• Prioritize optimizing full-order model interaction costs.

Topics

• Reduced-Order Models
• Incremental SVD
• Dynamical Systems
• Adaptive Modeling
• Computational Fluid Dynamics
• Nonlinear Problems

Code references

• raypretam/adaptive_llm_selection

Best for: AI Scientist, Research Scientist

Related on AIssential

• Physics-Guided Dimension Reduction for Simulation-Free Operator Learning of Stiff Differential--Algebraic Systems — An extended Newton implicit layer enables simulation-free, accurate operator learning for stiff DAEs by enforcing constraints and reducing dimensions.
• Conformal Prediction with Time-Series Data via Sequential Conformalized Density Regions — SCDR offers doubly robust, asymptotically valid conformal prediction for time-series, generating informative, potentially disconnected prediction sets.
• Online Bernstein-von Mises theorem — Sequential variational Bayesian updates can asymptotically match full posterior distributions with sufficient mini-batch sizes.
• Sub-JEPA: a simple fix to LeCun group's LeWorldModel that consistently improves performance [P] — Relaxing global latent space constraints with subspace regularization improves world model performance.
• Active Learning with Selective Time-Step Acquisition for PDEs — STAP reduces PDE surrogate model training costs by selectively acquiring only critical time steps.
• Scaling LLMs horizontally: hidden-state coupling without weight modification [R] — Residual Coupling enables horizontal scaling of frozen LLMs via linear bridges, enhancing performance without weight modification.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.