Justin Petucci, Penn State Institute for Computational and Data Sciences (ICDS) Research Innovations with Scientists and Engineers (RISE) artificial intelligence and machine learning team lead, leads and contributes to ongoing interdisciplinary research across Penn State in support of the University and ICDS' mission.
Petucci was first exposed to computational modeling and simulation using high-performance computing (HPC) while completing a master's degree in physics at Indiana University of Pennsylvania. He was a member of a group that received funding to build a small HPC cluster in the basement of the physics department. While at the University of Denver for his doctorate, he studied the absorption of gases, specifically greenhouse gases on different carbon nano materials using HPC.
It was after graduating that he moved back to State College and started working at ICDS' i-ASK help desk, which supports Penn State's largest HPC system. In 2019, Petucci joined the RISE team and became the AI/ML team lead in July of 2023.
Petucci's current ICDS-related work involves leveraging large language models (LLM) - an AI that can understand and process text - for HPC user support.
This project, which involves Simon Delattre, RISE engineer; Lindsay Wells, systems engineer; Emery Etter, research computing facilitation specialist; and Amit Amritkar, associate director for advanced computing, aims to use a retrieval-augmented generation (RAG) approach to create a chatbot that will answer questions for ICDS facility users. The RAG approach is a framework that aims to optimize LLM outputs.
"Simply put, RAG approaches take advantage of the zero-shot/in-context learning ability of modern LLMs by providing it with relevant context from a curated knowledge base to improve response quality to users," Petucci said. "If you give it [the model] a task that maybe it wasn't trained on but you provide sufficient information and instructions, a lot of times, the model can do a pretty good job at completing the task."
The researchers curated a local knowledge base comprising old ticket data, repeat topics and questions that come up at the help desk, as well as software user guides and manuals.
When users come to the chatbot with questions, their query is transformed into a numerical representation used to search the local knowledge base for relevant documents pertaining to the question. That information is then given to the language model to answer the initial question based on those identified resources.
"We are curating the knowledge base in hopes that it [the model] answers the questions more accurately," Petucci said. "There are many ways to tweak and improve RAG, but at its core, this is naive RAG. This is all within the context of the chatbot that we hope to expose to our users to answer general questions about our system or how to use or install software. If we can curate a good enough knowledge base and fine-tune on top of the naive RAG approach, this could be a good resource for our user base."
The research abstract led by Delattre, "Leveraging Large Language Models for HPC User Support: A RAG based Chatbot," was presented by Wells at the Practice and Expertise in Advanced Research Computing conference in July.
This team is also working to bring generative AI services to the ICDS Roar Collab HPC cluster. This is the newest cluster managed by the institute and was designed with collaboration in mind, according to the ICDS website. Roar Collab "allows for more frequent software updates and hardware upgrades to keep pace with researcher's changing needs."
Petucci and the research team said they hope to make open-source generative models and frameworks available to Roar Collab users, who will be able to test chatbots and not just generate text, but also images and video.
The team is also aiming to offer services around custom RAG-based chatbots, model fine-tuning and model deployment, Petucci said.
As part of his many collaborations, Petucci is a staff scientist at the Penn State Clinical and Translational Science Institute (CTSI) and works alongside Vasant Honavar, ICDS associate director, who leads the Informatics Core, on various projects.
Recently, CTSI seed grants were awarded to biomedical and clinical researchers to use AI and machine learning to improve health outcomes. Petucci is working on a variety of funded projects, one of which he leads with Monali Vasekar, associate professor of medicine at the Penn State College of Medicine.
This research team aims to apply deep learning AI algorithms to predict morbidities in patients with systemic cancer therapy induced pneumonitis using clinical and radiologic features.
"Cancer patients undergo different types of immunotherapies and chemotherapies," Petucci said. "In the end, these therapies can cause their own set of problems, and we plan to build a model that can be used for predictive purposes as well as gaining a better understanding of the underlying risk factors for various outcomes."
Researchers are leveraging multimodal electronic health record data, which includes radiological images. The team envisions using a trained model that can differentiate pneumonitis cases as part of a clinical decision support system (CDSS).
"Here, patterns or regions of interest on a patient's scan that are relevant to the underlying prediction would be highlighted to assist physicians in the early recognition of pneumonitis in equivocal cases," Petucci said. "That could potentially be powerful. Eliminating false positives and negatives is desirable. If we can make an accurate prediction, it could improve patient outcomes. Additionally, the CDSS tool could help augment the work of the radiologist by expediting the scan interpretation."
Petucci also works with Honavar and Ed O'Brien, professor of chemistry and ICDS co-hire, who are leading a $20 million U.S. National Science Foundation National Synthesis Center for Emergence in the Molecular and Cellular Sciences. The center is housed in the Huck Institutes of the Life Sciences at University Park.
