Research
Current
Role: PI (non-leading institution; Lead PI: Dr. Namgyun Kim at Texas A&M University)
Role: Co-PI (PI: Dr. Ali Shamshiripour)
Role: PI
Past
Role: PI
Role: PI
Role: PI
Role: Co-PI (PI: Dr. Chris Chaeha Lim)
Pacific Northwest National Laboratory
Publications
The complete, current list of journal articles is on Google Scholar.
Intellectual property
Software and data released by the lab. Everything here is citable and independently archived — each release carries its own DOI, so it can be referenced without going through the paper that describes it.
Software
LLM-based occupant agents grounded in American Time Use Survey microdata. Ships a Python agent library, a REST API, and an MCP server — with ATUS-informed activity scheduling across demographic groups, persona generation from EIA RECS appliance data, a memory stream, demand-response signal classification, and validation metrics for both behaviour and energy.
Generates representative U.S. residential occupancy schedules derived from smart thermostat data.
Datasets & benchmarks
120 commercial building knowledge graphs in Brick Schema v1.3, as RDF Turtle. Derived from DOE Commercial Prototype Building Models across eight building types and three ASHRAE code vintages, with sensor placement mined from 65 real buildings via Mortar. Ships 75 SPARQL queries with automatic scope detection; BuildGraph buildings reach a 91.5% Query Answerability Rate against that benchmark, versus 35.8% for public real-world Brick models.
6,136 executable SPARQL queries and 30,680 natural-language questions over 201 building knowledge graphs. Queries are validated by graph traversal before the natural-language questions are generated, across six query patterns and five vocabulary registers from ontology jargon to plain English. Includes a 173/28 building train–test split.
Eighteen building simulations from three DOE commercial prototypes (small, medium and large office, ASHRAE 90.1-2019) across three climate zones, each in fault-free and faulted variants. Combines Brick v1.3 knowledge graphs, 9.2 million hourly sensor readings across 1,050 streams, and 249 machine-gradable question–answer pairs. Faults are injected as physics parameters rather than post-hoc adjustments, so they propagate through the control system realistically.
Acknowledgements
With support from
In partnership with