BuildGraph: A Synthetic Multi-Archetype Building Knowledge Graph Dataset
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.
Tom: We're looking at the paper: "BuildGraph: A Synthetic Multi-Archetype Building Knowledge Graph Dataset".
Jane: Here's the gist of the paper. Semantic querying of building knowledge graphs (KGs) underpins the integration of artificial intelligence into building operations, from natural-language access to cross-building analytics BuildGraph is a synthetic building KG dataset of 120 buildings in the Brick Schema ontology, grounded in U.S.
Tom: Department of Energy prototype models and sensor-placement patterns from real buildings.
Jane: Dataset Scope and Grounding BuildGraph provides a collection of 120 synthetic building KGs in the Brick Schema ontology, derived from two empirical sources The dataset spans eight commercial building types across three ASHRAE energy-code vintages, with five realizations per archetype varying URI naming, sensor-attachment predicates, and topology.
Tom: Building configurations are derived from the U.S.
Paper summary: Jane: So we've covered the basics of what this paper is about. To wrap up the summary of "BuildGraph: A Synthetic Multi-Archetype Building Knowledge Graph Dataset," the main thesis is that building knowledge graphs are crucial for AI in building operations, but they are too scarce to be widely available because of proprietary issues.
Tom: Right. And BuildGraph addresses this scarcity by creating a synthetic dataset of one hundred twenty buildings in the Brick Schema ontology <ref:2610.00038#pg1,dataset of 120 buildings in the Brick Schema ontology>. It’s grounded in U.S. Department of Energy prototype models and sensor placement patterns from real buildings to make it as realistic as possible for research purposes, right?
Lu: Specifically, the dataset spans eight commercial building types across three ASHRAE energy-code vintages with five realizations per archetype that vary things like URI naming and sensor attachment predicates. This variety is key because it makes the data more challenging for testing.
Meng: And they used empirical data from the Mortar corpus to calibrate sensor placement, ensuring the frequency and type of sensors attached mirrors observed practice across real instrumented buildings. That’s an important detail for anyone building a model based on this.
Jane: It’s also important that they show structural completeness, achieving ninety-one point five percent Query Answerability Rate against thirty-five point eight percent for fifty-nine real-world Brick files, which shows the graphs are substantially more complete than what's currently public.
Tom: And beyond just completeness, they validated the dataset by reproducing real buildings’ sensor type proportions on their shared vocabulary with a cosine of zero point eighty six to zero point ninety four, which proves the data reflects real-world sensor usage patterns.
Lu: They also used SHACL schema conformance, and they achieved a one hundred percent conformance rate across all one hundred twenty buildings, confirming the present relationships are valid according to the schema <ref:2610.00038#pg1>. This gives us confidence in the structure they’ve built.
Meng: It’s clear this is a data creation effort aimed at solving a real bottleneck for AI integration into building operations where access to these knowledge graphs is currently blocked by proprietary restrictions and cost barriers.
Jane: Exactly, so they are providing a scalable, reproducible way to get high-quality, diverse building KGs for testing downstream tasks like text-to-SPARQL systems. This sets the stage for research that was previously limited by the small number of publicly available anonymized KG models.
Conclusion: Tom: So looking at the paper "BuildGraph: A Synthetic Multi-Archetype Building Knowledge Graph Dataset," it seems they’ve successfully created a large, structured synthetic building KG dataset for researchers to work with. The authors are Wooyoung Jung and his team.
Jane: It’s a very practical contribution because it directly tackles the problem of data scarcity in this area by providing something that is both diverse and grounded in real-world building models. They provide a documented, open-source methodology for generating these graphs which researchers can extend.
Lu: The main implication I see is that this dataset helps move text-to-SPARQL systems toward being better interfaces for facility managers by giving them the kind of structurally complete examples they need to train few-shot models effectively.
Meng: For me, the practical takeaway is that it gives us a way to generate diverse test cases tailored to specific portfolios, which is valuable for testing deployment stability on actual traffic.
Tom: So in simple terms, they’ve built a synthetic building KG dataset that helps researchers and engineers build better AI tools for accessing building knowledge through natural language queries without needing deep SPARQL expertise.
Jane: That’s the core idea: using synthetic data to solve a problem where real data is too scarce, and it gives us a solid foundation for the next wave of cross-building analytics in construction.
University of Arizona
cs.DB
Submitted: 2026-09-02
Updated: 2026-11-09
Comments: 25 pages, 3 figures. Data paper. Dataset, generator, and 75-query SPARQL benchmark openly available at https://github.com/humanbuildingsynergy/BuildGraph and https://doi.org/10.5281/zenodo.20015123. Re-set in the HUBS Lab paper layout (HUBS Lab Paper No. 2026-005); content unchanged
Code: https://github.com/humanbuildingsynergy/BuildGraph
License: http://creativecommons.org/licenses/by-nc-nd/4.0/
Importance score: 90/100
The gist: Semantic querying of building knowledge graphs (KGs) underpins the integration of artificial intelligence into building operations…
Key concepts
- Brick Schema Ontology
- This is the standardized vocabulary used to describe buildings in the knowledge graph. It provides a consistent way to label and connect different aspects of a building, such as equipment types, zones, and their relationships. This schema allows AI models to understand building data across various structures.
- Synthetic Building KG
- This is a dataset where buildings are artificially generated based on real-world prototypes. The researchers used DOE models and real sensor placement patterns to create these graphs. This method generates many varied building configurations, allowing researchers to test AI systems on diverse, realistic scenarios.
- QAR (Query Answerability Rate)
- QAR measures how well a knowledge graph can answer specific questions asked in natural language. It is calculated by checking the percentage of meaningful queries that successfully return at least one result. A high QAR indicates the graph is structurally complete and useful for complex information retrieval.
- Text-to-SPARQL Experiment
- This experiment tests how well a model can translate a natural language question into a precise SPARQL query that retrieves data from the building KG. It shows that few-shot prompting with examples significantly improves performance on cross-building generalization, proving the dataset's utility for complex analytical tasks.
Terminology
Summary
Semantic querying of building knowledge graphs (KGs) underpins the integration of artificial intelligence into building operations, from natural-language access to cross-building analytics BuildGraph is a synthetic building KG dataset of 120 buildings in the Brick Schema ontology, grounded in U. p. 0 (Corresponding author)
Abstract: Semantic querying of building knowledge graphs (KGs) underpins the integration of artificial intelligence
into building operations, from natural-language access to cross-building analytics, but such KGs are rarely public owing
to proprietary, security, and cost barriers. BuildGraph is a synthetic building KG dataset of 120 buildings in the Brick
Schema ontology, grounded in U. BuildGraph is a synthetic building KG dataset of 120 buildings in the Brick
Schema ontology, grounded in U. Department of Energy prototype models and sensor-placement patterns from real
buildings.Where the paper says it
Where the paper says it
Dataset Scope and Grounding
BuildGraph provides a collection of 120 synthetic building KGs in the Brick Schema ontology, derived from two empirical sources The dataset spans eight commercial building types across three ASHRAE energy-code vintages, with five realizations per archetype varying URI naming, sensor-attachment predicates, and topology. Building configurations are derived from the U. p. 1 This data paper introduces BuildGraph and its associated dataset: a collection of 120 synthetic building KGs in the
Brick Schema ontology, grounded in two independent empirical sources. First, building configurations are derived from the
U. First, building configurations are derived from the
U. Department of Energy (DOE) Commercial Prototype Building Models (U. Department of Energy (DOE) Commercial Prototype Building Models (U. Department of Energy 2024), published
specifications that define the HVAC system type, equipment counts, terminal unit configuration, and zone structure for eight
representative commercial building types spanning three energy code vintages. commercial building simulation research. Second, sensor
placement probabilities are calibrated against real buildings from the Mortar corpus (Fierro et al. Second, sensor
placement probabilities are calibrated against real buildings from the Mortar corpus (Fierro et al. 2019; Brick Schema
Community 2024), ensuring that the frequency and type of sensors attached to each equipment class reflects observed
practice across real instrumented buildings.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Structural Completeness Metrics
Dataset quality is assessed along two dimensions, first by a 75-query SPARQL benchmark measuring the Query Answerability Rate (QAR) BuildGraph achieves an overall QAR of 91.5%, compared to 35. p. 1 BuildGraph achieves an overall QAR of 91. 5%, compared to 35. 5%, compared to 35. 8% for a corpus of 59
publicly available real-world Brick building files – indicating that the generated graphs are substantially more structurally
complete with respect to this benchmark. 8% for a corpus of 59
publicly available real-world Brick building files – indicating that the generated graphs are substantially more structurally
complete with respect to this benchmark. Second, a downstream text-to-SPARQL experiment directly evaluates the dataset’s
utility for cross-building generalization. These omissions are invisible to SHACL because SHACL checks conformance
– whether what is present is schema-valid – not completeness – whether everything a query requires is there. The appropriate
measure is therefore query-centric: the Query Answerability Rate (QAR), defined as the fraction of applicable, semantically
meaningful SPARQL queries that return at least one result row when executed against a building KG.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Validation and Fidelity Checks
Dataset quality is validated across six complementary dimensions, including generation-time structural checks and SHACL schema conformance. The validator counts how many generated entity-property-entity triples match entries in the empirical pattern database, with a threshold of at least one floor-zone, zone-terminal, and AHU-terminal relationship required for structural failure. p. 2 SPARQL (Harris and Seaborne 2013) is
the W3C standard query language for RDF, enabling queries to pattern-match against building instance triples using Brick
class and property definitions to navigate relationships. A key capability is the property path operator (brick:feeds+),
which matches one or more consecutive brick:feeds edges, enabling a single query to trace the full supply chain from
chiller to zone. SAT-1
brick:hasPoint
SHACL: pass QAR: fail (no path to zone) SHACL: pass QAR: pass (path found)
(a) Schema-valid but structurally imcomplete (b) Schema-valid and structurally complete
Fig. SHACL conformance verifies that present relationships are schema-valid; QAR measures whether the graph
contains the relationships that queries require.Where the paper says it
Where the paper says it
Downstream Utility and Generalization
The text-to-SPARQL experiment with Gemma 4 (26B) reaches 27.3% Row-Matching F1 (+17. p. 0 A downstream text-to-SPARQL experiment with Gemma 4 (26B)
reaches 27. 3% Row-Matching F1 (+17. Downstream validation via a text-to-SPARQL evaluation using Gemma 4 (26B) (Gemma Team, Google
DeepMind 2026) on 12 held-out buildings, demonstrating that few-shot prompting with cross-building
BuildGraph text-to-SPARQL examples achieves 27. 3% Row-Matching F1 (the harmonic mean of precision
and recall over the set of result rows a query returns— measuring whether the query returns the correct answer
independent of SPARQL surface form, and defined in Section 6. Author keywords: building knowledge graph; synthetic dataset; Brick Schema; SPARQL; intelligent buildings; text-toSPARQL
Practical Applications
BuildGraph gives building practitioners a free, ready-made set of 120 realistic building knowledge graphs and a 75-question
query benchmark that would otherwise require proprietary, hard-to-obtain building data. Facility managers and controls
engineers can test fault-detection and energy-analytics queries across many building types and vintages before applying them
to their own sites; software developers can benchmark tools that translate plain-English questions into building-database
queries; and digital-twin developers can prototype and validate semantic applications without first assembling a real
building’s knowledge graph. 3% Row-Matching F1 (+17. 5 pp over zero-shot) on 12 held-out buildings. Real-world building KGs are rarely publicly
available due to proprietary restrictions, security concerns, and lack of instrumented buildings, limiting supply to a small
number of anonymized KG models released by national laboratories and universities. This scarcity is particularly acute
for text-to-SPARQL research – systems that translate natural language questions into executable SPARQL Protocol and
RDF Query Language (SPARQL) queries, emerging as a critical interface for facility managers and energy engineers to
access building knowledge without SPARQL expertise – because such systems require diverse, structurally complete KGs
as few-shot examples, training data, and evaluation benchmarks. 8% for a corpus of 59
publicly available real-world Brick building files – indicating that the generated graphs are substantially more structurally
complete with respect to this benchmark. Second, a downstream text-to-SPARQL experiment directly evaluates the dataset’s
utility for cross-building generalization. 3 Knowledge Graph Content
Entity Classes
The dataset uses 83 distinct Brick classes. Table 6 organizes the equipment entity classes by functional category.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Contribution and Extensibility
The primary contributions of this paper include a publicly accessible dataset of 120 synthetic building KGs in the Brick Schema ontology, spanning eight commercial building types, three ASHRAE energy code vintages, and various HVAC configurations. A documented, open-source methodology for reproducible synthetic building KG generation is provided to enable the research community to extend the dataset to additional building types or seeds. p. 1 A 75-query SPARQL benchmark for measuring structural quality of building KGs, with scope-detection
logic that automatically determines which queries are applicable to each graph and an evaluation framework
aligned with building complexity. A documented, open-source methodology for reproducible synthetic building KG generation, grounded in
published prototype models and empirical sensor placement data, enabling the research community to extend
the dataset to additional building types, ontology modes, or seeds. Facility managers and controls
engineers can test fault-detection and energy-analytics queries across many building types and vintages before applying them
to their own sites; software developers can benchmark tools that translate plain-English questions into building-database
queries; and digital-twin developers can prototype and validate semantic applications without first assembling a real
building’s knowledge graph. Because the generator is released, practitioners can also produce additional buildings tailored
to their own portfolio.Where the paper says it
Where the paper says it
Limitations and Future Directions
Known limitations include that the text-to-SPARQL experiment uses a structured natural-language summary rather than the raw RDF graph, which should be read as a conservative lower bound. The dataset does not include actual sensor measurement values, and its scope is limited to structural and topological information only. p. 7 “Scope” indicates which buildings include
instances of that category. Category Brick classes Scope
Building hierarchy Building, Floor, HVAC Zone All
Air-handling units AHU, Supply Fan, Return Fan,
Heating Coil, Cooling Coil,
Variable Frequency Drive,
Economizer Damper, Exhaust Damper,
Return Damper
All
Heat recovery Heat Exchanger (ERV) 2013 DOAS only
VAV terminals VAV, Variable Air Volume
Box With Reheat
VAV, PVAV, and
CAV systems
FCU terminals Fan Coil Unit FCU systems
Central plant Chiller, Cooling Tower, Boiler,
Chilled Water Pump, Hot Water Pump,
Condenser Water Pump
Central-plant buildings
Metering Electrical Meter 2004 and 2013
Sensor and Control Points
Sensor and control points span four domains, sampled from probabilities calibrated against the Mortar dataset. Structurally realistic synthetic KGs – matching
the equipment types, sensor placement patterns, topological relationships, and HVAC system configurations found in real
buildings – can substitute for real KGs as few-shot examples and evaluation benchmarks without privacy risk or access
barriers. The key design challenge is ensuring that the synthesis is empirically grounded rather than arbitrary: a randomly
generated graph can satisfy syntactic constraints while being physically implausible, producing a dataset that teaches a
model the wrong distribution. 0
Repository / DOI https://doi. 5281/zenodo. p. 6 0
Repository / DOI https://doi. 5281/zenodo. 5 pp over zero-shot) on 12 held-out buildings. BuildGraph gives facility managers
and digital-twin developers a testbed for portable analytics and natural-language interfaces, and its dataset, generator, and
benchmark are openly available at https://github. A high QAR confirms
that the A-box is structurally rich enough to support the full range of queries a downstream application would issue; a low
QAR reveals coverage gaps – analogous to the query-based completeness dimension in linked data quality literature (Zaveri
et al. 2016) – that would silently produce empty results in practice (Figure 2).Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
References
Almashor, M., Rana, M., McCulloch, J., Rahman, A., and Sethuvenkatraman, S. (2023). “What’s the point: AutoEncoding building point names. p. 4 (2021) confirmed that Brick offers the strongest foundation for semantic querying among existing
schemas. Complementary standards have since emerged – ASHRAE 223P (ASHRAE 2023) for connection-point topology,
the Building Topology Ontology (BOT) (Rasmussen et al. 2 Synthetic Building Data Generation
Prior work on generating building models falls into two categories distinct from semantic KG generation. Building
information modeling (BIM) tools produce models in the Industry Foundation Classes (IFC) format (buildingSMART
International 2013) capturing geometry and spatial hierarchy but not sensor attachments or typed semantic relationships;
energy simulation tools such as EnergyPlus (Crawley et al.Where the paper says it
Where the paper says it
ASHRAE (1995). p. 3 3 Related Work
3. 1 Building Metadata Standards and the Brick Schema
Early standardization efforts addressed communication interoperability (BACnet (ASHRAE 1995)) and tag-based operational
semantics (Project Haystack (Project Haystack Association 2014)), but lacked typed, named relationships between entities,
making SPARQL-based querying and ontological inference impossible (Bhattacharya et al. 3 Related Work
3. 1 Building Metadata Standards and the Brick Schema
Early standardization efforts addressed communication interoperability (BACnet (ASHRAE 1995)) and tag-based operational
semantics (Project Haystack (Project Haystack Association 2014)), but lacked typed, named relationships between entities,
making SPARQL-based querying and ontological inference impossible (Bhattacharya et al. 1 Introduction
Building knowledge graphs (KGs) are semantic models that represent the physical components, relationships, and
instrumentation of a building as a machine-readable graph. For example, a building KG encodes which air handling
units (AHUs) serve which heating, ventilation, and air conditioning (HVAC) zones, which sensors are attached to which
equipment, and how chilled water and hot water flow through the plant – expressed in a formal ontological vocabulary
such as the Brick Schema (Balaji et al. 1 Introduction
Building knowledge graphs (KGs) are semantic models that represent the physical components, relationships, and
instrumentation of a building as a machine-readable graph. For example, a building KG encodes which air handling
units (AHUs) serve which heating, ventilation, and air conditioning (HVAC) zones, which sensors are attached to which
equipment, and how chilled water and hot water flow through the plant – expressed in a formal ontological vocabulary
such as the Brick Schema (Balaji et al. 1 Introduction
Building knowledge graphs (KGs) are semantic models that represent the physical components, relationships, and
instrumentation of a building as a machine-readable graph. For example, a building KG encodes which air handling
units (AHUs) serve which heating, ventilation, and air conditioning (HVAC) zones, which sensors are attached to which
equipment, and how chilled water and hot water flow through the plant – expressed in a formal ontological vocabulary
such as the Brick Schema (Balaji et al.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
ASHRAE (2021). p. 4 (2021) confirmed that Brick offers the strongest foundation for semantic querying among existing
schemas. Complementary standards have since emerged – ASHRAE 223P (ASHRAE 2023) for connection-point topology,
the Building Topology Ontology (BOT) (Rasmussen et al.Where the paper says it
ASHRAE (2023). p. 4 (2021) confirmed that Brick offers the strongest foundation for semantic querying among existing
schemas. Complementary standards have since emerged – ASHRAE 223P (ASHRAE 2023) for connection-point topology,
the Building Topology Ontology (BOT) (Rasmussen et al. (2021) confirmed that Brick offers the strongest foundation for semantic querying among existing
schemas. Complementary standards have since emerged – ASHRAE 223P (ASHRAE 2023) for connection-point topology,
the Building Topology Ontology (BOT) (Rasmussen et al.Where the paper says it
Where the paper says it
Balaji, B., Bhattacharya, A. p. 1 2025), covers four buildings annotated in a mixture of Brick, ASHRAE 223P, and mixed-schema formats, providing
insufficient scale and structural diversity for systematic evaluation or cross-building generalization studies. This data scarcity problem is likely to persist for the foreseeable future due to a steep learning curve for practitioners to
learn, create, and maintain building KGs (Bhattacharya et al. 2025), covers four buildings annotated in a mixture of Brick, ASHRAE 223P, and mixed-schema formats, providing
insufficient scale and structural diversity for systematic evaluation or cross-building generalization studies. This data scarcity problem is likely to persist for the foreseeable future due to a steep learning curve for practitioners to
learn, create, and maintain building KGs (Bhattacharya et al. 2025), covers four buildings annotated in a mixture of Brick, ASHRAE 223P, and mixed-schema formats, providing
insufficient scale and structural diversity for systematic evaluation or cross-building generalization studies. This data scarcity problem is likely to persist for the foreseeable future due to a steep learning curve for practitioners to
learn, create, and maintain building KGs (Bhattacharya et al. 2025), covers four buildings annotated in a mixture of Brick, ASHRAE 223P, and mixed-schema formats, providing
insufficient scale and structural diversity for systematic evaluation or cross-building generalization studies. This data scarcity problem is likely to persist for the foreseeable future due to a steep learning curve for practitioners to
learn, create, and maintain building KGs (Bhattacharya et al. For example, a building KG encodes which air handling
units (AHUs) serve which heating, ventilation, and air conditioning (HVAC) zones, which sensors are attached to which
equipment, and how chilled water and hot water flow through the plant – expressed in a formal ontological vocabulary
such as the Brick Schema (Balaji et al. 2016; Fierro et al. For example, a building KG encodes which air handling
units (AHUs) serve which heating, ventilation, and air conditioning (HVAC) zones, which sensors are attached to which
equipment, and how chilled water and hot water flow through the plant – expressed in a formal ontological vocabulary
such as the Brick Schema (Balaji et al. 2016; Fierro et al. For example, a building KG encodes which air handling
units (AHUs) serve which heating, ventilation, and air conditioning (HVAC) zones, which sensors are attached to which
equipment, and how chilled water and hot water flow through the plant – expressed in a formal ontological vocabulary
such as the Brick Schema (Balaji et al. 2016; Fierro et al. For example, a building KG encodes which air handling
units (AHUs) serve which heating, ventilation, and air conditioning (HVAC) zones, which sensors are attached to which
equipment, and how chilled water and hot water flow through the plant – expressed in a formal ontological vocabulary
such as the Brick Schema (Balaji et al. 2016; Fierro et al. 2019), which published anonymized
Brick models for 45 real buildings from multiple U. university campuses and provided a unified SPARQL query
interface over building data. 2 Synthetic Building Data Generation
Prior work on generating building models falls into two categories distinct from semantic KG generation. Building
information modeling (BIM) tools produce models in the Industry Foundation Classes (IFC) format (buildingSMART
International 2013) capturing geometry and spatial hierarchy but not sensor attachments or typed semantic relationships;
energy simulation tools such as EnergyPlus (Crawley et al.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Balaji, B., Bhattacharya, A., Fierro, G., Gao, J., Gluck, J., Hong, D., Johansen, A., Koh, J., Ploennigs, J., Agarwal, Y., Berges, M., Culler, D. (2018). “Brick: Metadata schema for portable smart building applications. p. 4 Mortar showed that a single set of portable analytics applications – fault detection rules,
occupancy-based control algorithms, energy benchmarking queries – could be executed across heterogeneous buildings
without modification (Balaji et al. 2018), validating Brick’s core design goal. BuildGraph is a synthetic building KG dataset of 120 buildings in the Brick
Schema ontology, grounded in U. Department of Energy prototype models and sensor-placement patterns from real
buildings.Where the paper says it
Where the paper says it
Beckett, D. p. 5 2016; Fierro et al. 2022), serialized in RDF Turtle (TTL) format (Beckett et al. The Resource Description Framework
(RDF) (Cyganiak et al. 2014) provides this representation through a simple but expressive model: every fact is a triple of
the form subject–predicate–object. , brick:AHU as a subclass of brick:HVAC Equipment) and the formal semantics of
each relationship, while each building’s RDF file encodes its own instance data. SPARQL (Harris and Seaborne 2013) is
the W3C standard query language for RDF, enabling queries to pattern-match against building instance triples using Brick
class and property definitions to navigate relationships. , brick:AHU as a subclass of brick:HVAC Equipment) and the formal semantics of
each relationship, while each building’s RDF file encodes its own instance data. SPARQL (Harris and Seaborne 2013) is
the W3C standard query language for RDF, enabling queries to pattern-match against building instance triples using Brick
class and property definitions to navigate relationships. , brick:AHU as a subclass of brick:HVAC Equipment) and the formal semantics of
each relationship, while each building’s RDF file encodes its own instance data. SPARQL (Harris and Seaborne 2013) is
the W3C standard query language for RDF, enabling queries to pattern-match against building instance triples using Brick
class and property definitions to navigate relationships. , brick:AHU as a subclass of brick:HVAC Equipment) and the formal semantics of
each relationship, while each building’s RDF file encodes its own instance data. SPARQL (Harris and Seaborne 2013) is
the W3C standard query language for RDF, enabling queries to pattern-match against building instance triples using Brick
class and property definitions to navigate relationships.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Bhattacharya, A., Ploennigs, J., and Culler, D. (2015a). “Analyzing metadata schemas for buildings: The good, the bad, and the ugly. p. 4 2018), validating Brick’s core design goal. A subsequent review of 40 building metadata
schemas by Pritoni et al. 2 Synthetic Building Data Generation
Prior work on generating building models falls into two categories distinct from semantic KG generation. Building
information modeling (BIM) tools produce models in the Industry Foundation Classes (IFC) format (buildingSMART
International 2013) capturing geometry and spatial hierarchy but not sensor attachments or typed semantic relationships;
energy simulation tools such as EnergyPlus (Crawley et al. 2 Synthetic Building Data Generation
Prior work on generating building models falls into two categories distinct from semantic KG generation. Building
information modeling (BIM) tools produce models in the Industry Foundation Classes (IFC) format (buildingSMART
International 2013) capturing geometry and spatial hierarchy but not sensor attachments or typed semantic relationships;
energy simulation tools such as EnergyPlus (Crawley et al.Where the paper says it
Where the paper says it
Where the paper says it
Bhattacharya, A. A., Hong, D., Culler, D. E., Ortiz, J., Whitehouse, K., and Wu, E. (2015b). “Automated metadata construction to support portable building applications. p. 0 2016; Fierro et al. This structured representation underpins the integration
of artificial intelligence into building operations: it enables the semantic querying, inference, and reasoning behind
applications that range from natural-language interfaces—which let practitioners retrieve building information without
SPARQL expertise—to portable analytics that transfer across buildings.Where the paper says it
Brick Schema Community (2024). p. 1 Second, sensor
placement probabilities are calibrated against real buildings from the Mortar corpus (Fierro et al. 2019; Brick Schema
Community 2024), ensuring that the frequency and type of sensors attached to each equipment class reflects observed
practice across real instrumented buildings. The existing public building-KG resources (Table 1) each leave a different gap for method development. BuildingQA
pairs natural-language questions with SPARQL but covers only four buildings; Mortar publishes 45 real Brick models,
but as a fixed, uncontrolled collection with no paired query benchmark; and Open223 offers roughly nineteen illustrative
example buildings, likewise without one. The value of Brick at scale was demonstrated by the Mortar project (Fierro et al. 2019), which published anonymized
Brick models for 45 real buildings from multiple U.Where the paper says it
Where the paper says it
Where the paper says it
BrickSchema Community (2024). p. 1 Second, sensor
placement probabilities are calibrated against real buildings from the Mortar corpus (Fierro et al. 2019; Brick Schema
Community 2024), ensuring that the frequency and type of sensors attached to each equipment class reflects observed
practice across real instrumented buildings. 1 Introduction
Building knowledge graphs (KGs) are semantic models that represent the physical components, relationships, and
instrumentation of a building as a machine-readable graph. For example, a building KG encodes which air handling
units (AHUs) serve which heating, ventilation, and air conditioning (HVAC) zones, which sensors are attached to which
equipment, and how chilled water and hot water flow through the plant – expressed in a formal ontological vocabulary
such as the Brick Schema (Balaji et al. 00038v1 [cs. DB] 2 Sep 2026Where the paper says it
Where the paper says it
Where the paper says it
buildingSMART International (2013). p. 4 2 Synthetic Building Data Generation
Prior work on generating building models falls into two categories distinct from semantic KG generation. Building
information modeling (BIM) tools produce models in the Industry Foundation Classes (IFC) format (buildingSMART
International 2013) capturing geometry and spatial hierarchy but not sensor attachments or typed semantic relationships;
energy simulation tools such as EnergyPlus (Crawley et al. 2 Synthetic Building Data Generation
Prior work on generating building models falls into two categories distinct from semantic KG generation. Building
information modeling (BIM) tools produce models in the Industry Foundation Classes (IFC) format (buildingSMART
International 2013) capturing geometry and spatial hierarchy but not sensor attachments or typed semantic relationships;
energy simulation tools such as EnergyPlus (Crawley et al. 2 Synthetic Building Data Generation
Prior work on generating building models falls into two categories distinct from semantic KG generation. Building
information modeling (BIM) tools produce models in the Industry Foundation Classes (IFC) format (buildingSMART
International 2013) capturing geometry and spatial hierarchy but not sensor attachments or typed semantic relationships;
energy simulation tools such as EnergyPlus (Crawley et al.Where the paper says it
Where the paper says it
Where the paper says it
Crawley, D. p. 4 2 Synthetic Building Data Generation
Prior work on generating building models falls into two categories distinct from semantic KG generation. Building
information modeling (BIM) tools produce models in the Industry Foundation Classes (IFC) format (buildingSMART
International 2013) capturing geometry and spatial hierarchy but not sensor attachments or typed semantic relationships;
energy simulation tools such as EnergyPlus (Crawley et al. 2 Synthetic Building Data Generation
Prior work on generating building models falls into two categories distinct from semantic KG generation. Building
information modeling (BIM) tools produce models in the Industry Foundation Classes (IFC) format (buildingSMART
International 2013) capturing geometry and spatial hierarchy but not sensor attachments or typed semantic relationships;
energy simulation tools such as EnergyPlus (Crawley et al. BuildGraph is a synthetic building KG dataset of 120 buildings in the Brick
Schema ontology, grounded in U. Department of Energy prototype models and sensor-placement patterns from real
buildings. BuildGraph is a synthetic building KG dataset of 120 buildings in the Brick
Schema ontology, grounded in U. Department of Energy prototype models and sensor-placement patterns from real
buildings.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Cyganiak, R. p. 2 1 From Sensor Data to Queryable Knowledge Graphs
Integrating heterogeneous building sensor and equipment data across vendors and sites requires a common representation in
which entities, their types, and the relationships between them are explicitly encoded. The Resource Description Framework
(RDF) (Cyganiak et al. E-mail: wooyoung@arizona. (Corresponding author)
Abstract: Semantic querying of building knowledge graphs (KGs) underpins the integration of artificial intelligence
into building operations, from natural-language access to cross-building analytics, but such KGs are rarely public owing
to proprietary, security, and cost barriers.Where the paper says it
Where the paper says it
Dubey, M., Banerjee, D., Abdelkawi, A., and Lehmann, J. (2019). “LC-QuAD 2. p. 5 3 Text-to-SPARQL for Building Systems
Text-to-SPARQL has been studied extensively for general-purpose KGs, with large-scale benchmarks including Spider (Yu
et al. 2018), LC-QuAD (Trivedi et al. therefore constrained by the same data scarcity this work targets: a large, structurally consistent synthetic corpus offers one
route to the scale and variety such methods require (Nikolenko 2021). Synthetic data generation is well established in adjacent fields – from synthetic KGs and question-answering corpora to
simulated sensor streams – but building informatics has lacked an equivalent resource. 2 Synthetic Building Data Generation
Prior work on generating building models falls into two categories distinct from semantic KG generation. Building
information modeling (BIM) tools produce models in the Industry Foundation Classes (IFC) format (buildingSMART
International 2013) capturing geometry and spatial hierarchy but not sensor attachments or typed semantic relationships;
energy simulation tools such as EnergyPlus (Crawley et al.Where the paper says it
Where the paper says it
Where the paper says it
Fierro, G., Pritoni, M. p. 4 4 attempted to address this by layering Web Ontology Language (OWL) structure on the tag model, acknowledging
that the folksonomy approach was insufficient for programmatic reasoning (Pritoni et al. 2021) – yet retrofitting formal
semantics onto a flat tagging system proved architecturally incomplete. 4 attempted to address this by layering Web Ontology Language (OWL) structure on the tag model, acknowledging
that the folksonomy approach was insufficient for programmatic reasoning (Pritoni et al. 2021) – yet retrofitting formal
semantics onto a flat tagging system proved architecturally incomplete. 4 attempted to address this by layering Web Ontology Language (OWL) structure on the tag model, acknowledging
that the folksonomy approach was insufficient for programmatic reasoning (Pritoni et al. 2021) – yet retrofitting formal
semantics onto a flat tagging system proved architecturally incomplete. 4 attempted to address this by layering Web Ontology Language (OWL) structure on the tag model, acknowledging
that the folksonomy approach was insufficient for programmatic reasoning (Pritoni et al. 2021) – yet retrofitting formal
semantics onto a flat tagging system proved architecturally incomplete. It ships with a reusable 75-query
QAR benchmark that runs across all 120 buildings, and, because the generator is released, it is not a fixed corpus but an
open, extensible resource: users can create additional buildings—new archetypes, vintages, sizes, and naming or predicate
conventions—at any scale. To the author’s knowledge it is the first large-scale synthetic building-KG dataset released with
an accompanying evaluation protocol, giving research groups an identical, citable, and extensible substrate on which to
develop and reproducibly evaluate methods—the shared infrastructure that portable building-analytics research currently
lacks. BuildGraph is a synthetic building KG dataset of 120 buildings in the Brick
Schema ontology, grounded in U. Department of Energy prototype models and sensor-placement patterns from real
buildings.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Fierro, G., Pritoni, M., AbdelBaky, M., Lentzos, D., Prakash, A., Thomson, G., and Culler, D. E. (2022). “Brick schema: Building a unified metadata schema for buildings. p. 3 3 Related Work
3. 1 Building Metadata Standards and the Brick Schema
Early standardization efforts addressed communication interoperability (BACnet (ASHRAE 1995)) and tag-based operational
semantics (Project Haystack (Project Haystack Association 2014)), but lacked typed, named relationships between entities,
making SPARQL-based querying and ontological inference impossible (Bhattacharya et al.Where the paper says it
Gao, J., Ploennigs, J., and Berges, M. (2015). “A data-driven meta-data inference framework for building automation systems. p. 0 This structured representation underpins the integration
of artificial intelligence into building operations: it enables the semantic querying, inference, and reasoning behind
applications that range from natural-language interfaces—which let practitioners retrieve building information without
SPARQL expertise—to portable analytics that transfer across buildings. As the building industry increasingly embraces
semantic models, building KGs are becoming a critical data infrastructure, and their public availability is essential for
research progress. 2 Synthetic Building Data Generation
Prior work on generating building models falls into two categories distinct from semantic KG generation. Building
information modeling (BIM) tools produce models in the Industry Foundation Classes (IFC) format (buildingSMART
International 2013) capturing geometry and spatial hierarchy but not sensor attachments or typed semantic relationships;
energy simulation tools such as EnergyPlus (Crawley et al.Where the paper says it
Where the paper says it
Gemma Team, Google DeepMind (2026). p. 2 Downstream validation via a text-to-SPARQL evaluation using Gemma 4 (26B) (Gemma Team, Google
DeepMind 2026) on 12 held-out buildings, demonstrating that few-shot prompting with cross-building
BuildGraph text-to-SPARQL examples achieves 27. 3% Row-Matching F1 (the harmonic mean of precision
and recall over the set of result rows a query returns— measuring whether the query returns the correct answer
independent of SPARQL surface form, and defined in Section 6. 94),
evidence of realistic instrumentation where measurable. A downstream text-to-SPARQL experiment with Gemma 4 (26B)
reaches 27. As the building industry increasingly embraces
semantic models, building KGs are becoming a critical data infrastructure, and their public availability is essential for
research progress. 1
arXiv:2610. Downstream validation via a text-to-SPARQL evaluation using Gemma 4 (26B) (Gemma Team, Google
DeepMind 2026) on 12 held-out buildings, demonstrating that few-shot prompting with cross-building
BuildGraph text-to-SPARQL examples achieves 27. 3% Row-Matching F1 (the harmonic mean of precision
and recall over the set of result rows a query returns— measuring whether the query returns the correct answer
independent of SPARQL surface form, and defined in Section 6.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Hammar, K., Wallin, E. O., Karlberg, P., and Halleberg, D. (2019). “The RealEstateCore ontology. p. 4 Complementary standards have since emerged – ASHRAE 223P (ASHRAE 2023) for connection-point topology,
the Building Topology Ontology (BOT) (Rasmussen et al. 2021) for spatial decomposition within the W3C Linked Building
Data community, and RealEstateCore (Hammar et al. 2 Synthetic Building Data Generation
Prior work on generating building models falls into two categories distinct from semantic KG generation. Building
information modeling (BIM) tools produce models in the Industry Foundation Classes (IFC) format (buildingSMART
International 2013) capturing geometry and spatial hierarchy but not sensor attachments or typed semantic relationships;
energy simulation tools such as EnergyPlus (Crawley et al.Where the paper says it
Where the paper says it
Harris, S. and Seaborne, A. p. 2 , brick:AHU as a subclass of brick:HVAC Equipment) and the formal semantics of
each relationship, while each building’s RDF file encodes its own instance data. SPARQL (Harris and Seaborne 2013) is
the W3C standard query language for RDF, enabling queries to pattern-match against building instance triples using Brick
class and property definitions to navigate relationships. It spans eight commercial building types across three ASHRAE energy-code vintages, with five realizations per
archetype varying URI naming, sensor-attachment predicates, and topology. A 75-query SPARQL benchmark confirms
structural completeness: BuildGraph reaches 91. Real-world building KGs are rarely publicly
available due to proprietary restrictions, security concerns, and lack of instrumented buildings, limiting supply to a small
number of anonymized KG models released by national laboratories and universities. This scarcity is particularly acute
for text-to-SPARQL research – systems that translate natural language questions into executable SPARQL Protocol and
RDF Query Language (SPARQL) queries, emerging as a critical interface for facility managers and energy engineers to
access building knowledge without SPARQL expertise – because such systems require diverse, structurally complete KGs
as few-shot examples, training data, and evaluation benchmarks.Where the paper says it
Where the paper says it
Where the paper says it
Knublauch, H. p. 2 Query Answerability
Structural completeness cannot be measured by schema validation alone. The Shapes Constraint Language (SHACL) (Knublauch
and Kontokostas 2017) is the W3C standard for validating whether an RDF graph conforms to the property definitions
and cardinality constraints declared in the T-box: for example, that every brick:TemperatureSensor has the required
brick:isPointOf relationships to entities of type brick:Equipment (brick:isPointOf is the sensor-to-equipment
direction; its inverse, brick:hasPoint, attaches sensors from the equipment side). Query Answerability
Structural completeness cannot be measured by schema validation alone. The Shapes Constraint Language (SHACL) (Knublauch
and Kontokostas 2017) is the W3C standard for validating whether an RDF graph conforms to the property definitions
and cardinality constraints declared in the T-box: for example, that every brick:TemperatureSensor has the required
brick:isPointOf relationships to entities of type brick:Equipment (brick:isPointOf is the sensor-to-equipment
direction; its inverse, brick:hasPoint, attaches sensors from the equipment side). Query Answerability
Structural completeness cannot be measured by schema validation alone. The Shapes Constraint Language (SHACL) (Knublauch
and Kontokostas 2017) is the W3C standard for validating whether an RDF graph conforms to the property definitions
and cardinality constraints declared in the T-box: for example, that every brick:TemperatureSensor has the required
brick:isPointOf relationships to entities of type brick:Equipment (brick:isPointOf is the sensor-to-equipment
direction; its inverse, brick:hasPoint, attaches sensors from the equipment side).Where the paper says it
Where the paper says it
Where the paper says it
Koh, J., Balaji, B. p. 0 1 Introduction
Building knowledge graphs (KGs) are semantic models that represent the physical components, relationships, and
instrumentation of a building as a machine-readable graph. For example, a building KG encodes which air handling
units (AHUs) serve which heating, ventilation, and air conditioning (HVAC) zones, which sensors are attached to which
equipment, and how chilled water and hot water flow through the plant – expressed in a formal ontological vocabulary
such as the Brick Schema (Balaji et al. Author keywords: building knowledge graph; synthetic dataset; Brick Schema; SPARQL; intelligent buildings; text-toSPARQL
Practical Applications
BuildGraph gives building practitioners a free, ready-made set of 120 realistic building knowledge graphs and a 75-question
query benchmark that would otherwise require proprietary, hard-to-obtain building data. Facility managers and controls
engineers can test fault-detection and energy-analytics queries across many building types and vintages before applying them
to their own sites; software developers can benchmark tools that translate plain-English questions into building-database
queries; and digital-twin developers can prototype and validate semantic applications without first assembling a real
building’s knowledge graph. 2 Synthetic Building Data Generation
Prior work on generating building models falls into two categories distinct from semantic KG generation. Building
information modeling (BIM) tools produce models in the Industry Foundation Classes (IFC) format (buildingSMART
International 2013) capturing geometry and spatial hierarchy but not sensor attachments or typed semantic relationships;
energy simulation tools such as EnergyPlus (Crawley et al.Where the paper says it
Where the paper says it
Where the paper says it
Koh, J., Hong, D., Gupta, R., Whitehouse, K., Wang, H., and Agarwal, Y. (2018). “Plaster: An integration, benchmark, and development framework for metadata normalization methods. p. 4 A parallel line of work infers semantic metadata from existing building automation system (BAS) deployments via signal
processing and active learning (Gao et al. 2015), semi-automated metadata normalization (Koh et al. 2 Synthetic Building Data Generation
Prior work on generating building models falls into two categories distinct from semantic KG generation. Building
information modeling (BIM) tools produce models in the Industry Foundation Classes (IFC) format (buildingSMART
International 2013) capturing geometry and spatial hierarchy but not sensor attachments or typed semantic relationships;
energy simulation tools such as EnergyPlus (Crawley et al.Where the paper says it
Where the paper says it
Kovriguina, L. p. 5 2017; Dubey et al. 2019), and recent LLM-based approaches achieving high accuracy
on open-domain graphs (Kovriguina et al. Real-world building KGs are rarely publicly
available due to proprietary restrictions, security concerns, and lack of instrumented buildings, limiting supply to a small
number of anonymized KG models released by national laboratories and universities. This scarcity is particularly acute
for text-to-SPARQL research – systems that translate natural language questions into executable SPARQL Protocol and
RDF Query Language (SPARQL) queries, emerging as a critical interface for facility managers and energy engineers to
access building knowledge without SPARQL expertise – because such systems require diverse, structurally complete KGs
as few-shot examples, training data, and evaluation benchmarks.Where the paper says it
Where the paper says it
Lan, Y., He, G., Jiang, J., Jiang, J., Zhao, W. X., and Wen, J.-R. (2021). “A survey on complex knowledge base question answering: Methods, challenges and solutions. p. 5 therefore constrained by the same data scarcity this work targets: a large, structurally consistent synthetic corpus offers one
route to the scale and variety such methods require (Nikolenko 2021). Synthetic data generation is well established in adjacent fields – from synthetic KGs and question-answering corpora to
simulated sensor streams – but building informatics has lacked an equivalent resource. E-mail: wooyoung@arizona. (Corresponding author)
Abstract: Semantic querying of building knowledge graphs (KGs) underpins the integration of artificial intelligence
into building operations, from natural-language access to cross-building analytics, but such KGs are rarely public owing
to proprietary, security, and cost barriers.Where the paper says it
Where the paper says it
Mulayim, O. p. 1 This scarcity is particularly acute
for text-to-SPARQL research – systems that translate natural language questions into executable SPARQL Protocol and
RDF Query Language (SPARQL) queries, emerging as a critical interface for facility managers and energy engineers to
access building knowledge without SPARQL expertise – because such systems require diverse, structurally complete KGs
as few-shot examples, training data, and evaluation benchmarks. The only existing benchmark, BuildingQA (Mulayim
et al. The BuildingQA benchmark (Mulayim et al. 2025) is the only resource providing paired (natural
language, SPARQL) question-answer data, yet covers just 4 buildings. 2 Synthetic Building Data Generation
Prior work on generating building models falls into two categories distinct from semantic KG generation. Building
information modeling (BIM) tools produce models in the Industry Foundation Classes (IFC) format (buildingSMART
International 2013) capturing geometry and spatial hierarchy but not sensor attachments or typed semantic relationships;
energy simulation tools such as EnergyPlus (Crawley et al.Where the paper says it
Where the paper says it
Where the paper says it
Nikolenko, S. p. 1 Even if more real building KGs are released,
they will likely be limited in number, biased toward certain building types or instrumentation patterns, and encumbered by
proprietary restrictions that limit their utility for research and evaluation. Synthetic generation of building KGs is a promising strategy to address this data scarcity problem and provide a diverse,
scalable, and publicly accessible dataset for research (Nikolenko 2021). 2025), covers four buildings annotated in a mixture of Brick, ASHRAE 223P, and mixed-schema formats, providing
insufficient scale and structural diversity for systematic evaluation or cross-building generalization studies. This data scarcity problem is likely to persist for the foreseeable future due to a steep learning curve for practitioners to
learn, create, and maintain building KGs (Bhattacharya et al. com/humanbuildingsynergy/BuildGraph. Author keywords: building knowledge graph; synthetic dataset; Brick Schema; SPARQL; intelligent buildings; text-toSPARQL
Practical Applications
BuildGraph gives building practitioners a free, ready-made set of 120 realistic building knowledge graphs and a 75-question
query benchmark that would otherwise require proprietary, hard-to-obtain building data.Where the paper says it
Where the paper says it
Where the paper says it
Open223 Project (2023). p. 4 2019; Brick Schema Community 2024) is the largest of
these resources, offering publicly downloadable Brick models of 45 real buildings, but as a fixed collection of deployed real
models released without a paired query benchmark. The Open223 project (Open223 Project 2023) provides approximately
19 unique 223P example buildings from national laboratories, freely available but a small, illustrative set that likewise ships
no query benchmark. The existing public building-KG resources (Table 1) each leave a different gap for method development. BuildingQA
pairs natural-language questions with SPARQL but covers only four buildings; Mortar publishes 45 real Brick models,
but as a fixed, uncontrolled collection with no paired query benchmark; and Open223 offers roughly nineteen illustrative
example buildings, likewise without one.Where the paper says it
Where the paper says it
Pritoni, M., Paine, D., Fierro, G., Mosiman, C., Poplawski, M., Saha, A., Bender, J., and Granderson, J. (2021). “Metadata schemas and ontologies for building energy applications: A critical review and use case analysis. p. 4 2018), validating Brick’s core design goal. A subsequent review of 40 building metadata
schemas by Pritoni et al.Where the paper says it
Project Haystack Association (2014). p. 3 3 Related Work
3. 1 Building Metadata Standards and the Brick Schema
Early standardization efforts addressed communication interoperability (BACnet (ASHRAE 1995)) and tag-based operational
semantics (Project Haystack (Project Haystack Association 2014)), but lacked typed, named relationships between entities,
making SPARQL-based querying and ontological inference impossible (Bhattacharya et al. 3 Related Work
3. 1 Building Metadata Standards and the Brick Schema
Early standardization efforts addressed communication interoperability (BACnet (ASHRAE 1995)) and tag-based operational
semantics (Project Haystack (Project Haystack Association 2014)), but lacked typed, named relationships between entities,
making SPARQL-based querying and ontological inference impossible (Bhattacharya et al.Where the paper says it
Where the paper says it
Rasmussen, M. p. 4 (2021) confirmed that Brick offers the strongest foundation for semantic querying among existing
schemas. Complementary standards have since emerged – ASHRAE 223P (ASHRAE 2023) for connection-point topology,
the Building Topology Ontology (BOT) (Rasmussen et al. Complementary standards have since emerged – ASHRAE 223P (ASHRAE 2023) for connection-point topology,
the Building Topology Ontology (BOT) (Rasmussen et al. 2021) for spatial decomposition within the W3C Linked Building
Data community, and RealEstateCore (Hammar et al. 4 attempted to address this by layering Web Ontology Language (OWL) structure on the tag model, acknowledging
that the folksonomy approach was insufficient for programmatic reasoning (Pritoni et al. 2021) – yet retrofitting formal
semantics onto a flat tagging system proved architecturally incomplete. 4 attempted to address this by layering Web Ontology Language (OWL) structure on the tag model, acknowledging
that the folksonomy approach was insufficient for programmatic reasoning (Pritoni et al. 2021) – yet retrofitting formal
semantics onto a flat tagging system proved architecturally incomplete.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
RDFLib Contributors (2022). p. 2 3 (Fierro et al. 2022) introduced
SHACL shapes that enable automated conformance checking for any Brick model. Approximately 85% of sensors in each building carry a timeseries reference, reflecting realistic deployment
scenarios where some sensors may be offline or unregistered. 5 Methodology and Data Collection
The BuildGraph generation pipeline proceeds through five stages (Figure 3): (1) archetype definition – DOE prototype
parameters are encoded as Python dataclasses; (2) specification construction – each archetype is instantiated into a
BuildingSpec intermediate representation with sampled structural parameters; (3) graph building – the Brick builder
consumes the spec and produces an RDF graph with full topology, sensor attachments, and metadata; (4) serialization – the
graph is written to RDF Turtle; and (5) validation – generation-time structural checks and the full SPARQL benchmark
confirm structural completeness (Section 6).Where the paper says it
Where the paper says it
Trivedi, P. p. 5 3 Text-to-SPARQL for Building Systems
Text-to-SPARQL has been studied extensively for general-purpose KGs, with large-scale benchmarks including Spider (Yu
et al. 2018), LC-QuAD (Trivedi et al. 2018), LC-QuAD (Trivedi et al. 2017; Dubey et al. therefore constrained by the same data scarcity this work targets: a large, structurally consistent synthetic corpus offers one
route to the scale and variety such methods require (Nikolenko 2021). Synthetic data generation is well established in adjacent fields – from synthetic KGs and question-answering corpora to
simulated sensor streams – but building informatics has lacked an equivalent resource. 2 Synthetic Building Data Generation
Prior work on generating building models falls into two categories distinct from semantic KG generation. Building
information modeling (BIM) tools produce models in the Industry Foundation Classes (IFC) format (buildingSMART
International 2013) capturing geometry and spatial hierarchy but not sensor attachments or typed semantic relationships;
energy simulation tools such as EnergyPlus (Crawley et al.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
U.S. Department of Energy (2024). p. 1 Department of Energy (DOE) Commercial Prototype Building Models (U. Department of Energy 2024), published
specifications that define the HVAC system type, equipment counts, terminal unit configuration, and zone structure for eight
representative commercial building types spanning three energy code vintages. Department of Energy prototype models and sensor-placement patterns from real
buildings. It spans eight commercial building types across three ASHRAE energy-code vintages, with five realizations per
archetype varying URI naming, sensor-attachment predicates, and topology. The primary contributions of this paper are:
1. A publicly accessible dataset of 120 synthetic building KGs in the Brick Schema ontology, spanning eight
commercial building types (small office, medium office, large office, primary school, hospital, hotel, strip mall,
and outpatient healthcare), three ASHRAE energy code vintages, and various HVAC configurations; archetype
coverage validated against the Commercial Buildings Energy Consumption Survey (CBECS) 2018 data – a
national sample survey that details the stock of U.Where the paper says it
Where the paper says it
Where the paper says it
U.S. Energy Information Administration (2020). p. 4 2 Synthetic Building Data Generation
Prior work on generating building models falls into two categories distinct from semantic KG generation. Building
information modeling (BIM) tools produce models in the Industry Foundation Classes (IFC) format (buildingSMART
International 2013) capturing geometry and spatial hierarchy but not sensor attachments or typed semantic relationships;
energy simulation tools such as EnergyPlus (Crawley et al. The primary contributions of this paper are:
1. A publicly accessible dataset of 120 synthetic building KGs in the Brick Schema ontology, spanning eight
commercial building types (small office, medium office, large office, primary school, hospital, hotel, strip mall,
and outpatient healthcare), three ASHRAE energy code vintages, and various HVAC configurations; archetype
coverage validated against the Commercial Buildings Energy Consumption Survey (CBECS) 2018 data – a
national sample survey that details the stock of U. 2 Synthetic Building Data Generation
Prior work on generating building models falls into two categories distinct from semantic KG generation. Building
information modeling (BIM) tools produce models in the Industry Foundation Classes (IFC) format (buildingSMART
International 2013) capturing geometry and spatial hierarchy but not sensor attachments or typed semantic relationships;
energy simulation tools such as EnergyPlus (Crawley et al.Where the paper says it
Where the paper says it
Where the paper says it
Wang, Z., Sacks, R., and Yeung, T. (2022). “Exploring graph neural networks for semantic enrichment: Room type classification. p. 4 2018) – and all methods require existing
building data as input rather than generating KGs from scratch. Moreover, these learning-based approaches – including
graph-neural models for building semantic enrichment (Wang et al. 2001) produce thermodynamic outputs rather than RDF graphs. A parallel line of work infers semantic metadata from existing building automation system (BAS) deployments via signal
processing and active learning (Gao et al.Where the paper says it
Where the paper says it
Yu, T., Zhang, R., Yang, K., Yasunaga, M., Wang, D., Li, Z., Ma, J., Li, I., Yao, Q., Roman, S., Zhang. Z. and Radev. D. (2018). “Spider: A large-scale human-labeled dataset for complex and cross-domain semantic parsing and text-to-SQL task. p. 5 Synthetic data generation is well established in adjacent fields – from synthetic KGs and question-answering corpora to
simulated sensor streams – but building informatics has lacked an equivalent resource. 3 Text-to-SPARQL for Building Systems
Text-to-SPARQL has been studied extensively for general-purpose KGs, with large-scale benchmarks including Spider (Yu
et al. 2015b; Koh et al. 2018) – and all methods require existing
building data as input rather than generating KGs from scratch.Where the paper says it
Where the paper says it
Zaveri, A., Rula, A., Maurino, A., Pietrobon, R., Lehmann, J., and Auer, S. (2016). “Quality assessment for linked data: A survey. p. 3 The appropriate
measure is therefore query-centric: the Query Answerability Rate (QAR), defined as the fraction of applicable, semantically
meaningful SPARQL queries that return at least one result row when executed against a building KG. A high QAR confirms
that the A-box is structurally rich enough to support the full range of queries a downstream application would issue; a low
QAR reveals coverage gaps – analogous to the query-based completeness dimension in linked data quality literature (Zaveri
et al.Where the paper says it
Table 7
Table 7 reports SPARQL benchmark query categories.
Table 8 reports QAR ablation by pipeline stage. p. 8 Approximately 85% of sensors in each building carry a timeseries reference, reflecting realistic deployment
scenarios where some sensors may be offline or unregistered. 5 Methodology and Data Collection
The BuildGraph generation pipeline proceeds through five stages (Figure 3): (1) archetype definition – DOE prototype
parameters are encoded as Python dataclasses; (2) specification construction – each archetype is instantiated into a
BuildingSpec intermediate representation with sampled structural parameters; (3) graph building – the Brick builder
consumes the spec and produces an RDF graph with full topology, sensor attachments, and metadata; (4) serialization – the
graph is written to RDF Turtle; and (5) validation – generation-time structural checks and the full SPARQL benchmark
confirm structural completeness (Section 6).Where the paper says it
Table 9 reports CBECS coverage by BuildGraph archetype group. p. 6 20015123
Table 3. BuildGraph archetype coverage.Where the paper says it
Table 10 reports Sensor attachment rate comparison. p. 0 Department of Energy prototype models and sensor-placement patterns from real
buildings. It spans eight commercial building types across three ASHRAE energy-code vintages, with five realizations per
archetype varying URI naming, sensor-attachment predicates, and topology.Where the paper says it
Table 11 reports Sensor-composition fidelity against real Mortar buildings. p. 1 commercial building simulation research. Second, sensor
placement probabilities are calibrated against real buildings from the Mortar corpus (Fierro et al.Where the paper says it
Table 12 reports QAR by tier and query category for the full released dataset. p. 5 BuildGraph is built to fill this gap: it spans a balanced 8 × 3 × 5 factorial of eight building types, three
vintages, and five independent seeds (120 buildings), each released with its full provenance (archetype, vintage, seed) and
exact equipment and sensor inventory, and with naming scheme, attachment predicate, and topology varied across seeds by
design—the same heterogeneity real corpora exhibit, but here labeled and reproducible. It ships with a reusable 75-query
QAR benchmark that runs across all 120 buildings, and, because the generator is released, it is not a fixed corpus but an
open, extensible resource: users can create additional buildings—new archetypes, vintages, sizes, and naming or predicate
conventions—at any scale.Where the paper says it
Table 13 reports QAR comparison: 59-building Brick corpus and BuildGraph. p. 5 BuildGraph is built to fill this gap: it spans a balanced 8 × 3 × 5 factorial of eight building types, three
vintages, and five independent seeds (120 buildings), each released with its full provenance (archetype, vintage, seed) and
exact equipment and sensor inventory, and with naming scheme, attachment predicate, and topology varied across seeds by
design—the same heterogeneity real corpora exhibit, but here labeled and reproducible. It ships with a reusable 75-query
QAR benchmark that runs across all 120 buildings, and, because the generator is released, it is not a fixed corpus but an
open, extensible resource: users can create additional buildings—new archetypes, vintages, sizes, and naming or predicate
conventions—at any scale.Where the paper says it
Table 14 reports QAR ablation by pipeline stage. p. 8 Approximately 85% of sensors in each building carry a timeseries reference, reflecting realistic deployment
scenarios where some sensors may be offline or unregistered. 5 Methodology and Data Collection
The BuildGraph generation pipeline proceeds through five stages (Figure 3): (1) archetype definition – DOE prototype
parameters are encoded as Python dataclasses; (2) specification construction – each archetype is instantiated into a
BuildingSpec intermediate representation with sampled structural parameters; (3) graph building – the Brick builder
consumes the spec and produces an RDF graph with full topology, sensor attachments, and metadata; (4) serialization – the
graph is written to RDF Turtle; and (5) validation – generation-time structural checks and the full SPARQL benchmark
confirm structural completeness (Section 6).Where the paper says it
Table 15 reports Text-to-SPARQL Row-Matching F1 by query category. p. 2 Downstream validation via a text-to-SPARQL evaluation using Gemma 4 (26B) (Gemma Team, Google
DeepMind 2026) on 12 held-out buildings, demonstrating that few-shot prompting with cross-building
BuildGraph text-to-SPARQL examples achieves 27. 3% Row-Matching F1 (the harmonic mean of precision
and recall over the set of result rows a query returns— measuring whether the query returns the correct answer
independent of SPARQL surface form, and defined in Section 6.Where the paper says it
Overall QAR
Overall QAR 87.5% 96.9% 92.9% 96.5% 91.5%.
Category Overall QAR 87. p. 1 First, a 75-query SPARQL benchmark measures the Query
Answerability Rate (QAR) – the fraction of applicable queries returning at least one result row – across five query categories
and four building complexity tiers. BuildGraph achieves an overall QAR of 91.Where the paper says it
Category T1 T2 T3 T4 Overall QAR 87. p. 1 First, a 75-query SPARQL benchmark measures the Query
Answerability Rate (QAR) – the fraction of applicable queries returning at least one result row – across five query categories
and four building complexity tiers. BuildGraph achieves an overall QAR of 91.Where the paper says it
Category Zero-shot QAR 9. p. 0 3% Row-Matching F1 (+17. 5 pp over zero-shot) on 12 held-out buildings. 3 Knowledge Graph Content
Entity Classes
The dataset uses 83 distinct Brick classes. Table 6 organizes the equipment entity classes by functional category.Where the paper says it
Where the paper says it
Overall QAR 9. p. 1 First, a 75-query SPARQL benchmark measures the Query
Answerability Rate (QAR) – the fraction of applicable queries returning at least one result row – across five query categories
and four building complexity tiers. BuildGraph achieves an overall QAR of 91. First, a 75-query SPARQL benchmark measures the Query
Answerability Rate (QAR) – the fraction of applicable queries returning at least one result row – across five query categories
and four building complexity tiers. BuildGraph achieves an overall QAR of 91.Where the paper says it
Where the paper says it
Category 5 Zero-shot QAR 0. p. 0 3% Row-Matching F1 (+17. 5 pp over zero-shot) on 12 held-out buildings. Nodes (ovals) are building entities; labeled arrows are typed semantic
relationships. Following the two brick:feeds arrows in sequence – from AHU-1 through VAV-3 to Zone-B – is an
example of a multi-hop query.Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
- Data Scarcity Mitigation via Synthetic Generation
The system can be improved by using BuildGraph as a testbed for portable analytics and natural-language interfaces
to bridge the gap caused by proprietary data scarcity. This allows facility managers and digital-twin developers to test fault-detection and energy-analytics queries across many building types and vintages before applying them to their own sites.
- Enhanced Text-to-SPARQL Capability
The AI system can be improved by using BuildGraph examples in a few-shot setting, as demonstrated by Gemma 4 achieving 27.3% Row-Matching F1 (+17.5 pp over zero-shot)
on held-out buildings. This capability allows the system to translate plain-English questions into building-database queries
with improved accuracy than zero-shot methods alone.
- Structural Completeness Guarantee
The system can be improved by incorporating generation invariants from the BuildGraph pipeline, which includes checks like Every floor contains at least one zone (brick:hasPart)
and Every AHU reaches at least one brick:HVAC Zone via brick:feeds+
. This ensures that the data used for downstream tasks is structurally complete
and reduces failures caused by missing relationships.
- Archetype Agnostic Analytics
The system can be improved by leveraging the dataset's cross-building nature, allowing a query written for one building to be executed across the full corpus without modification.
This enables systematic evaluation of how well an analytics application generalizes across diverse building types and vintages.
- Domain Model Development Foundation
The system can be improved by using the structured metadata, such as triples like meta:buildingType records the archetype identifier combining type and vintage,
to support any model that reasons about HVAC systems, equipment relationships, or sensor coverage.
This provides a foundation for developing domain-specific language and reasoning models.
Abstract
Semantic querying of building knowledge graphs (KGs) underpins the integration of artificial intelligence into building operations, from natural-language access to cross-building analytics, but such KGs are rarely public owing to proprietary, security, and cost barriers. BuildGraph is a synthetic building KG dataset of 120 buildings in the Brick Schema ontology, grounded in U.S. Department of Energy prototype models and sensor-placement patterns from real buildings. It spans eight commercial building types across three ASHRAE energy-code vintages, with five realizations per archetype varying URI naming, sensor-attachment predicates, and topology. A 75-query SPARQL benchmark confirms structural completeness: BuildGraph reaches 91.5% Query Answerability Rate versus 35.8% for 59 real-world Brick files. Independently, it reproduces real buildings' sensor-type proportions on their shared vocabulary (cosine 0.86-0.94), evidence of realistic instrumentation where measurable. A downstream text-to-SPARQL experiment with Gemma 4 (26B) reaches 27.3% Row-Matching F1 (+17.5 pp over zero-shot) on 12 held-out buildings. BuildGraph gives facility managers and digital-twin developers a testbed for portable analytics and natural-language interfaces, and its dataset, generator, and benchmark are openly available at https://github.com/humanbuildingsynergy/BuildGraph.
Sources
Related papers
- Beyond Overlap: Hierarchical Graph Abstraction for Out-of-Core Vector Search
- EMA: Approximate Nearest Neighbor Search with General Attribute Filtering and Dynamic Updates
- Open Biomedical Knowledge Graphs at Scale: Construction, Federation, and AI Agent Access with Samyama Graph Database
- Samyama: A Unified Graph-Vector Database with In-Database Optimization, Agentic Enrichment, and Hardware Acceleration
- GenTUS: Table Union Search via Constrained Generation of Semantic Table Identifiers
- Common-Neighbor-Count-Based Representative Possible World Finding on Uncertain Graphs