AI 中文总结
针对II型超新星样本分析的计算瓶颈，提出基于STELLA的两个神经网络替代模型，利用自动编码器、模拟器及正则化等技术，应用于多个超新星，能快速进行物理特征描述并大幅减少贝叶斯推断时间。
AI 中文摘要
为解决分析来自时空遗产调查等II型超新星样本的计算瓶颈，我们提出了两个基于STELLA的神经网络替代模型：一个用于低能量爆炸且可能与星际物质（CSM）相互作用的相互作用模型，以及一个用于标准无相互作用IIP型超新星的光球模型。每个模型都使用自动编码器压缩光谱能量分布，并使用模拟器将物理参数映射到潜在空间。潜在混合正则化改善了潜在空间的连续性，相互作用模型使用ResNet块，光球模型使用2D CNN。它们在测试集上的归一化重建MSE分别约为9.1e - 5和1.0e - 4。将模型应用于SN 2005cs、SN 2012aw和SN 1999em，得到了相应的恒星初始质量等结果，这些替代模型将完整贝叶斯推断从数天减少到数分钟，能够快速对大型超新星样本进行物理特征描述。
英文摘要
To address the computational bottleneck of analyzing type II supernova samples from surveys such as the Legacy Survey of Space and Time, we present two STELLA-based neural-network surrogates: an interaction model for low-energy explosions with possible circumstellar-material (CSM) interaction and a photospheric model for standard interaction-free SNe IIP. Each uses an autoencoder to compress spectral energy distributions and an emulator to map physical parameters to the latent space. Latent-mixup regularization improves latent-space continuity, with ResNet blocks used for the interaction model and 2D CNNs for the photospheric model. Their normalized test-set reconstruction MSEs are approximately 9.1e-5 and 1.0e-4, respectively. Applied to SN 2005cs, the interaction model favors a low-mass progenitor, M_ZAMS = 10.40(+0.04/-0.05) M_sun, and confined dense CSM, providing a scenario consistent with direct imaging and helping resolve the historical mass discrepancy. For SN 2012aw, it recovers M_ZAMS = 11.05(+0.06/-0.06) M_sun, consistent with previous studies. For SN 1999em, the photospheric model gives M_ZAMS = 10.05(+0.07/-0.04) M_sun, broadly consistent with preexplosion imaging limits without explicit CSM modeling. These surrogates reduce full Bayesian inference from days to minutes and enable rapid physical characterization of large supernova samples.
Comments29 pages, 7 figures; accepted for publication in Physical Review D
