Li Wang
Li Wang is a machine learning researcher and engineer based at the University of Western Australia. He works at the intersection of AI and science, building systems that help researchers make sense of large, complex datasets — with a particular focus on radio astronomy.
He holds a PhD in Computer Science from the University of Technology Sydney and has spent several years at CSIRO developing and deploying ML solutions on high-performance computing infrastructure. His work spans the full stack — from model design and training to scalable production pipelines.
Beyond astronomy, he has collaborated across disciplines including environmental engineering and biomedical systems, driven by a belief that the same underlying methods can unlock insight in almost any data-rich domain.
He is broadly interested in where modern ML is heading — foundation models, efficient inference, and what it means to build AI that is genuinely useful in scientific practice.
Session
The next generation of radio astronomy telescopes, including the Square Kilometre Array (SKA), will generate unprecedented data volumes, requiring highly automated and reliable approaches for source finding and catalogue construction. Existing neutral hydrogen (HI) source-finding methods, including numerical algorithms and machine-learning approaches, have demonstrated strong performance but often rely on survey-specific preprocessing, parameter tuning, or architectures that do not fully exploit the three-dimensional structure of spectral-line data. We present the HI Segment Anything Model (HISAM), an automated image-to-catalogue pipeline for HI source finding based on a pretrained three-dimensional large vision model. HISAM adapts a volumetric segmentation architecture to HI spectral-line cubes through domain-specific modifications and fine-tuning, and integrates source segmentation with catalogue construction and false-positive rejection to transform input image cubes into science-ready source catalogues. We evaluate HISAM on the SKA Science Data Challenge 2 simulation. HISAM achieves competitive completeness and reliability for sources with sufficiently strong signals, while its overall completeness decreases when evaluated across the full source population.