This paper presents a comprehensive framework for quantifying neural network waveform errors and incorporating them into gravitational-wave parameter estimation. A multi-fidelity approach decomposes waveform uncertainty into epistemic and aleatoric components, enabling rigorous propagation of surrogate errors through the likelihood function. The method combines deep ensemble uncertainty estimation with physics-informed constraints to produce well-calibrated error bounds. A novel likelihood modification marginalizes over waveform uncertainty, ensuring conservative parameter posteriors.
Key findings
Developed a taxonomy of uncertainty sources in neural network waveform surrogates.
Proposed a multi-fidelity uncertainty quantification framework combining deep ensembles and physics-informed constraints.
Derived a modified likelihood function that marginalizes over waveform uncertainty.
Validated the framework through extensive simulations on binary black hole waveforms.
Limitations & open questions
The framework's computational cost may negate some benefits of neural surrogates.
Surrogate uncertainty estimates require further validation against ground-truth errors from high-fidelity simulations.