Jing Lu

As the lead of a research project, I developed a deep-learning emulator to replace computationally expensive radiative-transfer simulations used to infer physical properties from observed supernova spectra. I built a configuration-driven workflow spanning training-data generation and processing, DenseNet-based emulator training, model validation, and integration into a broader inference pipeline. The resulting emulator accelerated spectral synthesis by roughly 100 million times while maintaining approximately 1% prediction error, enabling efficient exploration of previously impractical high-dimensional parameter spaces. I also leveraged HPC resources for generating large-scale simulation datasets and MPI to accelerate the inference process.

As the lead of a research project, I developed data-driven models to learn and reconstruct time-varying spectral patterns from observational datasets. I designed the full modeling workflow, including data preprocessing, dimensionality reduction with PCA, and Gaussian Process regression (with a conditional variational autoencoder as an alternative method) to generate continuous spectral templates. The main method is packaged into an open-source Python package (BYOST). I evaluated the models against existing templates and quantified their impact on downstream analysis. The resulting framework reduced a key source of systematic error by approximately 90% relative to previous templates.