RP: Multi-Fidelity Surrogates for Gappy, Expensive Simulation Data in Gravitational-Wave Astronomy
Proceedings of the AAAI Symposium Series · 06.11.2026 · 10 (1) · 619-625
DOI: 10.1609/aaaiss.v10i1.43464 ↗Kurz gefasst
Fachartikel in der Zeitschrift Proceedings of the AAAI Symposium Series, online erschienen am 06.11.2026. Verfasst von Tousif Islam.
Abstract (Originalsprache)
Resource-constrained surrogate learning must often work with tens to hundreds of expensive simulations whose lengths and accuracies vary substantially. Requiring complete trajectories discards useful data; attaching approximate outputs to short simulations introduces additional modeling error. We propose a multi-fidelity construction in which a few long, accurate trajectories provide the reduced basis and shorter, heterogeneous trajectories constrain its coefficient maps across parameter space. We fit these maps directly on observed windows through weighted, regularized linear regression, without completing each training trajectory. The surrogate then predicts full trajectories at new parameters. We demonstrate the method on gravitational-wave datasets with 45 and 174 numerical trajectories for map training and validation, and assess optional approximate-model enrichment separately. Retaining partial numerical trajectories reduces the two-parameter median held-out error on 87 nonzero-spin examples from 1.2e-3 to 6.2e-5 at fixed approximate-model augmentation. Direct fitting also improves on coefficient-target regression by factors of 3.8-6.4 in the one-parameter task. These results show how unequal-duration simulations can support compact surrogates without a large collection of complete training examples.
Zitieren
APA
Islam, T. (2026). RP: Multi-Fidelity Surrogates for Gappy, Expensive Simulation Data in Gravitational-Wave Astronomy. Proceedings of the AAAI Symposium Series, 10(1), 619-625. https://doi.org/10.1609/aaaiss.v10i1.43464
Vancouver
Islam T. RP: Multi-Fidelity Surrogates for Gappy, Expensive Simulation Data in Gravitational-Wave Astronomy. Proceedings of the AAAI Symposium Series. 2026;10(1):619-625. doi:10.1609/aaaiss.v10i1.43464
BibTeX
@article{Islam2026,
author = {Islam, Tousif},
title = {RP: Multi-Fidelity Surrogates for Gappy, Expensive Simulation Data in Gravitational-Wave Astronomy},
journal = {Proceedings of the AAAI Symposium Series},
year = {2026},
volume = {10},
number = {1},
pages = {619--625},
doi = {10.1609/aaaiss.v10i1.43464}
}
EUVIMED