AI for Operations

AI for Operations (ORNL/DOE)

An umbrella initiative consisting of several application-focused research projects to build AI capabilities for internal ORNL/DOE day-to-day operations. Supported by an initial base funding from the Office of the Deputy Director, Laboratory Operations, ORNL since 2024.

Hazardous Event Forecasting at Department of Energy Sites

Team: Tirthankar Ghosal (PI),  Prasanna Balaprakash (Co-PI), Dana Hewit, Ryan Burchfield, Lauren Deck, Sanjay Das (ECO-S Intern from UTD), Ran Elgedawy (GRO Intern from UTK), Ethan Seefried (ECO-S Intern from CSU), Gavin Wiggins

Develop an Information Retrieval-based Generative AI Tool to "forecast" hazardous events and perform predictive analytics. Generate reports to identify vulnerabilities in Work Plans and learn from "Lessons Learned" from past events and accidents. Text mining 30 years of DOE event database, OSHA reports, ORNL ACTS, SBMS, Training records, RSS, Work Plan system to enhance safety procedures in ORNL and DOE labs. Several sub-tasks contribute to the main task, e.g., classification of event types into multiple event categories based on the textual description, failure mode analysis, causal analysis, etc.

High-Risk Property Prediction

Team: Tirthankar Ghosal (PI), Silvers Saffel, Lauren Deck, Vanessa Lama, Maria Mahbub, Brian Starks, Christopher Polcheck, Sanjay Das (ECO-S Intern from UTD)

Develop a secure and trustworthy AI to identify and track 'High-Risk Property' items. Reason over classification and develop a feedback loop to learn over time. The developed system needs to stay updated with the latest changes in DOE policies and revise the decision accordingly. Team includes people from NCCS, ITSD, and HRP SMEs. Draft HRP letters and notify stakeholders in a human-in-the-loop automated fashion. Develop an on-premise, robust, and current HRP reasoning system that learns from human feedback.

https://ornl.sharepoint.com/Pages/Article.aspx?articleId=46721 

Multimodal Information Retrieval and Complex Engineering Diagram Parsing

Team: Tirthankar Ghosal (PI), Ethan Seefried (ECO-S Intern from CSU), Rob Saethre (Spallation Neutron Source, ORNL), Prasanna Balaprakash (CSMD)

This is a proof-of-concept project sponsored by the Neutron Science Directorate (NScD) to perform semantic information search and retrieval over multimodal SNS documents (texts and engineering diagrams ~ 750k). The project involves the parsing of complex engineering diagrams for semantic multimodal metadata extraction using Computer Vision and NLP techniques. One of the goals from this project is to develop datasets, models, and artifacts and open up the problem to the larger community to co-develop the solution.

GenAI-based S&T Matrix Classification

Team: Tirthankar Ghosal (PI), Scott Ewing, Layla Marshall, Christopher Stahl, Jessica Brownfield, Prasanna Balaprakash, Shaun Gleason

The S&T Matrix was created to address concerns regarding the protection of sensitive technologies and research areas that could impact US national security and economic competitiveness. The project involves building a 'human-in-the-loop' AI-infrastructure to identify projects, papers, and documents that may contain sensitive technological information, and the level of sensitivity with appropriate reasoning on the outcome.

GenAI for Activity-Based Work Control - Chemistry and Complex Predictive Chemistry

Team: Tirthankar Ghosal (PI), Kimberley Jeskie, Chrissi Schnell, Susan Fiscor, Ran Elgedawy (GRO Intern from UTK), Ryan Burchfield, Subhamay Pramanik

This project is a spin-off from the main 'Hazardous Event Forecasting', but for chemistry. As ORNL moves towards Activity-Based Work Control (ABWC), this project explores how AI can complement the incoming ABWC system by reducing processing and generation time by X fold, performing efficient retrieval of 'critical information' from trusted sources, and making the system 'conversational'. Another vision of this project is to employ AI to do 'complex predictive chemistry' for predicting chemical-related hazards in workplaces.

Artificial Curator

Team: Tirthankar Ghosal (PI), Olga Kuchar, Meghan Berry

To develop an Artificial Curator for Constellation involving semantic metadata extraction and generation from scientific artifacts (papers, datasheets, etc.). The project will resume operation upon subsequent funding.