Mega AI

Inspired by recent successes in artificial intelligence (AI)-driven discovery and multiple breakthroughs in growing the size, accuracy, and generalizability of AI models, Mega AI—an internal PNNL investment—seeks to develop massive-scale, self-supervised, multimodal foundation models of scientific knowledge capable of general-purpose inferences to enable reasoning with existing knowledge and discovery of new knowledge across science and security domains.

Mega AI

Donald Jorgensen | Pacific Northwest National Laboratory 

Mega AI aims to deliver cutting-edge, next-generation AI capabilities unique to the Department of Energy national lab complex. Mega AI addresses existing research gaps in narrow AI, specifically in scaling AI to hundreds of billions of parameters, multimodal representation learning, multitask general-purpose inferences, the need for increased generalizability and rapid development and deployment of AI technologies, and the usability and assurance of AI models for science and security applications.

  • General-purpose AI: Mega AI develops foundational models of scientific knowledge to augment scientists with the advanced ability to perceive and reason at a scale previously unimagined. Unlike narrow AI models, a massive-scale foundation model (aka neural platform) is a single model that learns from huge amounts of raw unlabeled data across multiple modalities and is multi-purpose. Therefore, it can be rapidly adapted to a wide range of useful tasks, such as knowledge summarization, information extraction, hypothesis generation and validation, question answering, classification, recommendation etc.
  • Scalable AI: Mega AI operates on the cloud at the scale of hundreds of millions (hundreds of gigabytes) of scholarly documents: papers, patents, reports, books, and lines of code.
  • Multimodal AI: Mega AI focuses on many modalities including but not limited to text, images (plots and figures), structured data (tables, references, graphs), equations, code, and domain-specific modalities.
  • Trusted and Responsible AI: Mega AI evaluates large-scale foundation models of scientific knowledge across many in- and out-of-domain benchmarks and downstream tasks, and on interactive evaluation tools to assure model accountability, robustness, fairness, and transparency.
  • Human-AI Reasoning: Mega AI delivers augmented intelligence solutions driven by a human-centered partnership model of scientist and AI working together to enhance cognitive performance, including learning (knowing, remembering), decision-making (judging), and reasoning.

Mega AI Focus Areas: Climate Security

Climate security encompasses the physical, economic, or societal changes associated with climate change that imminently and substantially alter the political stability or degree of human security in a country or region; the national security of the United States; the military, political, or economic interests of allies and partners; or citizens of the United States abroad.

  • Mega AI focuses on generating new knowledge summaries that incorporate a broader knowledge base than can be done by individual scientists or teams of scientists.
  • Mega AI's massive-scale, cross-domain, knowledge-based approach aims to generate domain-specific scientific hypotheses that are novel and testable.

Mega AI Focus Areas: Molecular Chemistry

Chemistry and materials science literature holds massive amounts of multimodal knowledge encoded as text, tables, and images about experimental property measurements, synthesis procedures and methods, organic reaction pathways, and more.

  • Mega AI focuses on encoding multimodal knowledge representations of experimental design, system conditions, and synthesis procedures to augment scientists’ reasoning with this knowledge at a much larger scale than is currently possible.
  • Mega AI aims to learn joint representations of knowledge extracted from literature and databases of molecular structures and properties to make predictions to discover novel synthesis mechanisms, reveal previously unknown molecular functionalities, and design new molecular structures in silico.