Research

Azton Wells's AstroMLab research series explores the development of highly specialized AI assistants for astronomy. Beginning with AstroMLab 1's "Astronomy Jeopardy" to assess baseline AI knowledge, the work progressed to advanced model development. AstroMLab 3 introduced an 8B-parameter large language model, achieving GPT-4o level performance in astronomical tasks through domain-specific training. This was further expanded in AstroMLab 4, which deployed a 70B-parameter domain-specialized reasoning model. This larger model demonstrated benchmark-topping performance in astronomy Q&A, significantly enhancing the accuracy and utility of AI systems for complex scientific inquiry. The core methodology emphasizes deep domain specialization to achieve superior performance in astronomical knowledge retrieval and reasoning.

Learn more about this research →

Azton Wells's research advances AI methodologies to accelerate and validate scientific discovery. HiPerRAG introduces a high-performance Retrieval Augmented Generation (RAG) framework designed to efficiently synthesize vast scientific literature, enabling the extraction of novel insights and hypothesis generation. This method directly addresses information overload, accelerating knowledge synthesis. Complementarily, EAIRA establishes a crucial methodology for rigorously evaluating AI models functioning as scientific research assistants. This framework ensures the reliability, accuracy, and trustworthiness of AI contributions, providing a standardized approach to assess their utility and validity in research contexts. Together, these works aim to build robust, verifiable AI tools that enhance data interpretation and drive scientific progress.

Learn more about this research →

Azton Wells' research illuminates early universe galaxy evolution, particularly the profound impact of primordial stars and their supernovae. Utilizing advanced numerical simulations, including the Phoenix suite, the work meticulously models the chemical enrichment of dark matter minihalos at redshifts $z \gtrsim 10$. Key methods involve simulating gas accretion, cooling, and virialization, alongside the heterogeneous dispersal of metals from the first supernovae. This enrichment is shown to be crucial for facilitating second-generation star formation. The research highlights how both internal and external enrichment mechanisms shape the environments conducive to the birth of the first galaxies, bridging the gap between primordial star-forming regions and subsequent galaxy assembly.

Learn more about this research →

Azton Wells' research exemplifies the powerful application of machine learning in astrophysics, notably through "Predicting Localized Primordial Star Formation with Deep Convolutional Neural Networks." This work employs Deep Convolutional Neural Networks (DCNNs) to analyze intricate cosmological simulations. The core method involves training these DCNNs to accurately identify and predict specific regions conducive to primordial star formation, a critical process in the early universe. By leveraging DCNNs, Wells significantly enhances the efficiency and precision of analyzing vast astrophysical datasets. This pioneering approach offers a robust tool for extracting subtle, localized patterns from complex data, ultimately accelerating discovery and refining our understanding of the universe's initial structure formation.

Learn more about this research →