Deep MUSE observations of SNR 0509-67.5 reveal a double degenerate merger progenitor

Deep MUSE observations of SNR 0509-67.5 reveal that the coronal [Fe xiv] λ5303 emission line appears with either one or two velocity components across the entire remnant, arising from reverse-shocked ejecta moving toward and away from the observer. A supervised dense neural network classifies each spaxel and fits Gaussian profiles plus a linear function to the observed line emission. We measure a bulk Doppler velocity of −1000 ± 60 km s−1, interpreted as the line-of-sight component of the primary white dwarf’s orbital velocity in a double-degenerate merger. The red- and blue-shifted ejecta map shows a flattened edge along the north-eastern rim, consistent with the companion’s shadow, indicating a binary companion was present at explosion. Modelling this feature as a cone anchored at the explosion centre and applying Bayesian inference, we recover the cone’s orientation and half-opening angle. We then use the Eggleton Roche-lobe relation to infer properties of the companion. The companion was likely a ∼0.6 M⊙ white dwarf with radius ∼9800 km and orbital velocity ∼1700 km s−1 at the time of explosion. Together, these results provide a complete dynamical picture of a Type Ia supernova progenitor system whose maximum-light spectrum is independently constrained by light echo observations.

Improvements to AI Systems:

1. Physics-informed neural network architecture for multi-component spectral line fitting:

• Replace the generic dense neural network with a hybrid model that explicitly encodes the expected physics (e.g., Gaussian line profiles, linear continuum, known emission-line ratios) as inductive biases in the loss function.

• The improved system can simultaneously classify spaxels and fit multiple velocity components with uncertainty quantification, reducing false positives in low signal-to-noise regions.

2. Bayesian inference with domain-specific priors for geometric parameter recovery:

• Integrate a probabilistic programming layer that uses the Eggleton Roche-lobe relation and supernova explosion geometry as prior distributions, rather than uninformed priors.

• The improved system can automatically infer cone orientation, half-opening angle, and companion mass/radius/orbital velocity with credible intervals, even when the data are sparse or asymmetric.

3. Self-supervised representation learning for velocity map segmentation:

• Use a contrastive learning pretraining step on synthetic supernova remnant velocity fields (generated from hydrodynamical simulations) to learn robust features of edge-on and shadowed ejecta.

• The improved system can detect subtle morphological features (e.g., flattened edges, shadow cones) in new remnants without manual feature engineering, and can generalize to other Type Ia remnants.

4. Uncertainty-aware multi-objective optimization for Doppler velocity extraction:

• Combine the neural network’s classification output with a differentiable Bayesian optimizer that jointly minimizes spectral fitting residuals and maximizes physical consistency (e.g., momentum conservation across red/blue components).

• The improved system can produce a global 3D kinematic map of the remnant, including line-of-sight velocity gradients, and flag spaxels where the single- or double-Gaussian assumption breaks down.

5. Transfer learning for light-echo-constrained progenitor classification:

• Fine-tune the trained network on synthetic spectra from double-degenerate merger models (varying companion mass, separation, explosion delay) to create a classifier that predicts the most likely progenitor configuration from observed maximum-light spectra.

• The improved system can automatically rank candidate companion properties (mass, radius, orbital velocity) for any Type Ia supernova with light-echo data, enabling rapid population studies.

What the improved AI system can do:

• Automatically produce a full 3D kinematic reconstruction of a supernova remnant from integral-field spectroscopy, including error bars on every velocity component.

• Infer the progenitor system’s binary parameters (companion mass, radius, orbital velocity, and orientation) directly from the observed ejecta geometry, without manual modeling.

• Distinguish between single-degenerate and double-degenerate scenarios with quantified confidence, using only the shape of the [Fe xiv] emission and the shadow cone.

• Predict the companion’s properties for new remnants in minutes, enabling large-scale surveys of Type Ia progenitors across different galaxies.

Deep MUSE observations of SNR 0509-67.5 reveal that the coronal [Fe, xiv] lambda 5303 emission line appears with either one or two velocity components across the entire remnant, arising from reverse-shocked ejecta moving toward and away from the observer. A supervised dense neural network classifies each spaxel and fits Gaussian profiles plus a linear function to the observed line emission. We measure a bulk Doppler velocity of-1000 plus or minus60 km s-1, interpreted as the line-of-sight component of the primary white dwarf's orbital velocity in a double-degenerate merger. The red- and blue-shifted ejecta map shows a flattened edge along the north-eastern rim, consistent with the companion's shadow, indicating a binary companion was present at explosion. Modelling this feature as a cone anchored at the explosion centre and applying Bayesian inference, we recover the cone's orientation and half-opening angle. We then use the Eggleton Roche-lobe relation to infer properties of the companion. The companion was likely a about 0.6 M white dwarf with radius about 9800 km and orbital velocity about 1700 km s-1 at the time of explosion. Together, these results provide a complete dynamical picture of a Type Ia supernova progenitor system whose maximum-light spectrum is independently constrained by light echo observations.