We present a novel deep learning framework, AIAI (Artificial Intelligence-based Abundance Inference), for reconstructing the chemical and density structures of Type Ia supernovae (SNe Ia) from synthetic and observational spectra. Using a multiresidual neural network (MRNN) trained on TARDIS-generated spectral models, we predict elemental abundances with high accuracy and apply the method to a sample of 22 observed SNe Ia near B-band maximum light. The model successfully captures correlations between \(^{56}\) Ni abundance and light curve parameters such as stretch and \(\Delta m_{15}\) and reveals the time evolution of \(^{56}\) Ni and \(^{56}\) Co in the ejecta. While predicted B-band magnitudes are systematically brighter than those derived from light curve fitting, this discrepancy is attributed to limitations in the TARDIS radiative transfer assumptions. We also apply our method to construct delay time distributions (DTDs) for SNe Ia using spatially resolved star formation histories from 96 host galaxies observed with VLT/MUSE. A power-law DTD model yields best-fit parameters of a characteristic delay time \(\tau = 120^{+142}{-83}\) Myr and a decay slope \(s = -1.41^{+0.32}_{-0.33}\) , consistent with previous studies and favoring a double-degenerate progenitor scenario. Our results highlight the potential of combining physically motivated simulations with neural networks to constrain SN explosion physics and progenitor demographics. The methods developed here offer a promising path forward for improving SN Ia cosmology and understanding systematic uncertainties in distance measurements, especially at high redshifts.