GURU is an autonomous system that drives expert workflows
in software. In order to guarantee reliability and avoid hallucination,

MSBAI employs hierarchical planning. We’ve built a scalable

approach to preparing every skills agent we train, for each

additional software package we register on the system.

GURU is an
autonomous system that drives expert workflows in software. In order to guarantee reliability and avoid hallucination, MSBAI employs
hierarchical planning. We’ve built a scalable approach to preparing every skills agent we train, for each additional software package we register on the system.

hierarchical planning

We start from the cheapest, most plentiful data to minimize the need for expensive manual user sessions, or running massive numbers of simulations. This ensures the system performs expert work reliably, using a training strategy that maximizes the speed of adding more modular knowledge.

Hierarchical
Learning

Monitoring a small number of human expert sessions for Reinforcement Learning from Human Feedback (RLHF) to build the adaptive plan generated from navigation and action selection, and construct Ontologies.
Training the system how to perform actions by analyzing manuals and (video) tutorials. The planning structure includes transformer World Models, Joint Embeddings networks for interpreting user requests, learning from heterogeneous data (text, images, video) and synthesizing new language based commands, and world models for encoding dynamics and improving generalizability.
Graph Neural Networks (GNNs) navigate Graphical User Interfaces (GUIs), databases and templates with computer vision, providing the foundation for higher-level actions.

Research papers

Digital Engineering

Digital Transformation

HPC

Deep Learning

Reinforcement Learning

Artificial Intelligence

Simulation Workflows in Minutes, at Scale for Next-Generation HPC

Author

Allan Grosvenor, Anton Zemlyansky, Dwyer Deighan, and Dustin Sysko

Author

As the White House pushes to decarbonize energy, the Department of Energy (DOE)’s National Reactor Innovation Center (NRIC) has an urgent need to decrease the cost and schedule for new reactor design and construction in support of the Advanced Construction Technology (ACT) initiative. Current lead time for new reactors is 20–30 years and costs $10–$15 billion. This must be dramatically reduced to bring advanced reactors online. Digital Engineering, leveraging the best multiphysics simulation and high-performance computing (HPC), offers us a unique opportunity to lead these efforts, but a paradigm shift in engineering is mandatory: right now on the order of only 1% of engineers regularly use simulation as a tool in their design toolbox—meaning it is unusual for engineers to create virtual prototypes and broadly explore the available space of design options, and test and evolve them with modeling and simulation.

Massive virtual prototype explorations are rarely done in new product development, because engineering modeling & simulation packages take months-to-years to learn, and setup of a new simulation can often require hours of laborious work. We must enable a new user to set up and run thousands of models quickly to evolve virtual prototypes. DOE has spent nearly $100 million in taxpayer funds, and decades of development, to advance HPC. There is massive untapped potential in the thousands of simulation packages in existence, and the commercial cloud computing that is plentiful and affordable today. Computational physics and HPC needs to be put in the hands of every engineer to begin a renaissance in construction and manufacturing. We present an autonomous system built to hyper-enable engineers, and the work we’ve conducted using the Summit supercomputer to pursue it.