Exploring the origins of high-velocity features in SNe Ia with the spectral synthesis code TARDIS

TL;DR

This study tackles the origin of high-velocity features (HVFs) in Type Ia supernovae by combining 1D radiative-transfer modelling with Gaussian density enhancements in the outer ejecta. Using TARDIS to generate photospheric (PV) spectra, the authors train neural-network emulators on a four-parameter grid (amplitude, centroid, width, and Si abundance) of outer-density boosts and apply MCMC to jointly fit HVF evolution across six well-sampled SNe Ia. They find that a single density enhancement can reproduce silicon HVFs well but cannot simultaneously reproduce calcium HVFs, suggesting that HVFs arise from more complex ejecta structures and potentially multiple line-forming regions. The results imply that neither the delayed-detonation nor the double-detonation explosion models alone can explain HVFs, underscoring the need for higher-fidelity NLTE physics, 3D geometry, and broader observational coverage, especially at very early times. Overall, the work highlights the importance of detailed HVF modelling, early-time spectroscopy, and multi-species constraints to advance our understanding of the outer ejecta in SNe Ia.

Abstract

Appearing as secondary higher-velocity absorption components, high-velocity features (HVFs) have been observed in several absorption lines in many Type Ia supernovae (SNe Ia). The frequency and ubiquity of these components in silicon and calcium features specifically indicates that the mechanism through which they form must be common occurrence among the majority of SNe Ia. Here we present modelling of the HVF evolution in a sample of six well observed SNe Ia with the radiative-transfer code tardis. A base model is derived for each of the SNe to reproduce the photospheric velocity components, followed by a grid of simulations with Gaussian enhancements to the density profile at high velocities. We train a set of neural networks to emulate the impact of these density enhancements upon the simulated silicon line profile. These networks are subsequently used within a Markov-Chain Monte Carlo (MCMC) framework to infer the density enhancement parameters that most closely reproduce the HVF evolution. While we obtain good matches for the silicon profile, we find that a single density enhancement alone cannot simultaneously produce the observed silicon and calcium HVF evolution. Our findings indicate that neither the delayed-detonation mechanism, nor the double-detonation mechanism can produce these HVFs, suggesting that something may be missing from the models.

Figures (16)

• Figure 1: Spectral sequences of the six SNe Ia to be modelled. Fits to the Siii $\lambda6355$ feature are displayed in black for each spectrum. In the case of a preferred two-component fit, the individual components are displayed in grey. The spectra are plotted in normalised luminosity and offset, as is the case with all spectra presented throughout this study.
• Figure 2: Species densities for the PV models for the six SNe Ia in our sample. These species densities are calculated as the product of the abundance and density profiles and can be used to compare models with different density profiles. The faint red and faint blue profiles correspond to the HESMA delayed-detonation ddt_models and double-detonation double_det_models models, respectively. The colours of the PV model profiles correspond to the colours in Fig. \ref{['fig:spectral_series']}.
• Figure 3: Schematic view of the chosen NN architecture. The blue nodes on the left correspond to the four inputs governing the density enhancement and silicon abundance, which are fed into the input layer. The input layer and hidden layer are both comprised of 200 neurons with a softplus activation function and are represented by the grey nodes. Finally the orange nodes correspond to the normalised luminosity outputs at the 75 wavelength points across the range of the Siii $\lambda6355$ feature.
• Figure 4: Performance summary of the 27 NNs (four or five per SN based on the number of observed spectral epochs). The evolution of the MSE is shown against the training epoch in the top panels. The middle and bottom panels show the median and worst case predictions, respectively, from the simulated test data corresponding to each SN epoch. The grey lines are the unsmoothed TARDIS outputs and the colours lines are the NN predictions. The median and worst predictions are not necessarily at the same epoch for each SN.
• Figure 5: Best fitting density enhancement model spectra (colour solid) compared to observed spectra (grey) and the Siii $\lambda6355$ region from the PV simulations (colour dotted). The faint regions of the model spectra correspond to the regions which are not constrained in the MCMC fitting. The asterisk by the $-15.7$ d spectrum of SN 2018cnw indicates that this spectrum was not used in the MCMC parameter inference (see Section \ref{['sec:method:neural_networks:mcmc']}).