演講資訊

摘要：

Type Ia supernovae are the best-calibrated distance indicators in cosmology, yet we still do not know which stars explode or how they burn. They carry the local value of the Hubble constant and the recent hints of evolving dark energy, so their physics has become part of the cosmological question. In this talk I show how machine learning and AI agents are changing what can be done with the supernova data we already have. First, AI agents have re-reduced more than two decades of archival VLT spectropolarimetry end to end, producing a uniform archive whose spectra are anchored to the published photometry. Spectropolarimetry reveals the three-dimensional shape of the ejecta. Across the absorption lines the polarization often traces a loop, a twist with depth that may record the spin of the exploding white dwarf. Yet the population as a whole shows no preferred handedness. Second, I introduce SNAIL-2, a transformer model that learns the spectrophotometric manifold of SNe Ia from the ultraviolet to the near-infrared, trained on spectra and photometry together. From one or two spectra it predicts a supernova's full spectral evolution. I illustrate this with recent nearby supernovae and with the gravitationally lensed SN H0pe at z = 1.78. Finally, I return to cosmology. Combined with the Twins Embedding, SNAIL-2 lets JWST use SNe Ia beyond z ~ 1.6 to measure the matter density Omega_M almost free of the dark-energy degeneracy. That would be a clean test of the current disagreement between supernovae and baryon acoustic oscillations.

This event is supported by the University Academic Alliance in Taiwan (UAAT).