FusionSci: Augmented Intelligence to Empower Interdisciplinary Research

The exponential growth of scientific knowledge makes cross-disciplinary collaboration extremely difficult, limiting knowledge exchange, researcher productivity, and slowing scientific discovery. The team at Aptima, along with research partner Pacific Northwest National Laboratory (PNNL) and partners at National Renewable Energy Laboratory (NREL) and Jefferson Laboratory propose to develop FusionSciŚan artificial intelligence- (AI) powered augmented intelligence capability leveraging large language models (LLMs, e.g., LLaMa, Claude 2) and multimodal foundation models (e.g., SciTune, GPT4) with a corpus of millions of scientific documents to empower cross-disciplinary scientific collaborations. FusionSci will use techniques such as prompt engineering, in- context learning, and fine-tuning with curriculum learning to synthesize and reason with scientific knowledge including declarative, procedural, and conditional across disciplines. In Phase I, FusionSci will focus on developing an interactive tool using human-centered design. Our experiments will integrate LLMs and foundation models into a cloud-based generative AI testbed and benchmark performance on scientific QA, summarization, and search tasks. To ensure that FusionSci models are producing high-quality, accurate, and relevant scientifically-valid outputs, we will implement a critique model that utilizes the tree of thought and graph of thought prompting with step-wise verification. Finally, we will perform studies with domain scientists to evaluate FusionSci usability and generalizability across science domains, measuring team productivity, composition, and dynamics, then incorporating end-user feedback to improve FusionSci model performance. If successful, FusionSci will synthesize insights across disciplines by processing vast diverse data at scale previously unimagined, enabling innovative solutions; summarize and extract specialized knowledge, saving researchers time and increasing their productivity; make complex science accessible to the public, promoting engagement; and democratize knowledge by making specialized information accessible to all.