ALCHEM LabOntology
Materials, design strategies, computations, property values and findings extracted from our papers, connected in an ontology-based knowledge graph — ask in plain language.
Answers are based on information extracted from the papers by a large language model (LLM). Lab members are reviewing the extracted information, which may be incomplete or inaccurate. Check the cited page in the paper or supporting information.
Loading…
What is an ontology?
In information science, an ontology is an explicit description of a field: the kinds of things in it (classes), how they are related (relations) and what they are called (vocabularies), written so that people and computers read it the same way. The word comes from philosophy, where ontology is the study of what exists; in computing it is often defined, after Gruber (1993), as “an explicit specification of a conceptualization”.
More about ontologies
Stated in the paper
Re-SZIO after CP / post-OER Re-SZIOMaterial systemderived_from Re-SZIOMaterial systemderived_from SZIO (Zr:Ir = 1:2)Material system
Inferred by the reasoner
Re-SZIO after CP / post-OER Re-SZIOMaterial systemdescends_from SZIO (Zr:Ir = 1:2)Material system
derived_from is a sub-property of descends_from, and descends_from is transitive. So from the two stated facts the OWL reasoner concludes that the catalyst after the stability test also descends from SZIO, a link the paper never writes as such. Search uses it: these links are marked “inferred”.
[73] Re-SZIO for acidic OER · This paper’s facts →
Not just keywords. A keyword list names things; a taxonomy also says what is a kind of what (HER is an electrochemical reaction); a database schema fixes how records are stored. An ontology also says what the relations mean, so facts from different papers written with the same classes and relations can be joined and queried together, and a reasoner can infer facts that were never written down.
Axioms. What makes this possible are the ontology’s axioms: statements that always hold, such as “a material is never a reaction”, “a value belongs to one material”, “a value computed with a simulation is not a measured value”, or “a Faradaic efficiency lies between 0 and 100 %”. A reasoner checks every extracted fact against them. In our graph they found values filed under the wrong property, units printed wrongly in a paper, and a value marked as measured although it names its simulation; values that break one are kept out of rankings and charts, and marked “!” in the tables.
Inference. The reasoner adds facts from the axioms. “Made from” and “part of” carry over, so a question about PS-b-PDMS also finds the zinc-anode host (Zn@CPC-Co) made from it through several steps, and shows the path. A concept covers those under it (electrocatalysis → HER, OER, ORR). A computational model is tied to the real sample the paper builds it for, so a material’s computed and measured values meet. Whether a paper reports calculations, experiments or both is inferred too. Such results are marked “inferred”. The axioms and hierarchy were drafted by one AI model, reviewed by a second from another company, and decided by the PI.
In science. The Gene Ontology in biology and ChEBI for chemical entities are long-standing examples. For materials, EMMO (Elementary Multiperspective Material Ontology) is being developed as a common framework for materials modelling and characterisation. Ontologies are usually written in the W3C standards RDF and OWL; ours can be downloaded in OWL below.
Ontology
From papers to a graph of facts
An ontology fixes what can be said about a paper — its classes, their relations and vocabularies — so that facts from different papers can be compared, joined and queried.
Classes and relations. A paper reports material systems (composition, structure, defects, dopants), the design strategies applied to them, how they were made and computed, the reactions they serve, the values measured or calculated, and the findings that tie these together. The same kind of material in different papers is linked as one shared material kind, never as the same sample.
Read from the full papers, with their sources. An LLM reads each paper in full and fills the ontology through a fixed schema; every fact carries the page it comes from and the run that wrote it, and a fact whose quoted text cannot be found in the paper is dropped. Who did what is recorded per research group (the lab or a partner group), not per person: two AI models read the paper on their own and a third decides where they differ.
Asked, linked and checked. During search, an LLM, when available and permitted, is used only to translate your question into a query using the ontology’s vocabulary; otherwise the site uses keyword search. Answers are assembled from the papers and information found in the graph. The vocabulary is linked to Wikidata, and each release of the graph is tested with competency questions, SHACL shapes and OWL rules. These test its structure; extraction and question interpretation can still be wrong, so check the sources.
Schema
How our data is written
The classes of the ALCHEM ontology and the relations between them, with the number of records of each: a paper’s bibliographic side, who did what (per research group, not per person), and what was extracted from it, with the material kinds shared across papers. The schema is written in LinkML in three modules, from which the OWL ontology, SHACL shapes and the extraction format are generated.
Linked toWikidata (vocabulary terms, SKOS mappings)CRediT contributor rolesDOI · ORCID · ROR (see also)
Knowledge graph v1 (ontology + data)Ontology (OWL)Data (TriG, named graphs with provenance)SHACLCompetency questionsDCATREADMECC BY 4.0
Knowledge graph
Which terms meet in the papers
One view of the knowledge graph: the most frequent reactions, material classes, design strategies and computational methods across 76 papers, linked when they are used in the same papers (thicker the more papers). A link means they appear together, not a scientific relation. Choose a term for its papers and its Wikidata item.
Our data
What is in the graph
- Papers
- 76
- Papers read in full
- 76
- Material systems
- 2,101
- Design strategies
- 363
- Computations
- 171
- Reaction studies
- 221
- Property values
- 11,075
- Findings
- 1,882
- Authors
- 372
- Institutions
- 79
- Shared material kinds (each in 2 or more papers)
- 76
- Papers with who did what, per research group
- 76
- Papers whose calculations the lab did
- 72
- Vocabulary terms linked to Wikidata
- 108
- Grants acknowledged (distinct numbers)
- 154
Extracted with gpt-6.1-sol (OpenAI) from the full texts of the papers, including supporting information. The extraction model and prompt were selected by comparing their outputs against a gold set of five papers. Lab members are reviewing the extracted results.
Checks Each release is tested with 13 competency questions (SPARQL, with expected and forbidden answers), SHACL shapes (with deliberately broken records that must fail) and OWL 2 RL rules (no contradictions found). These test the graph’s structure, not that every extracted fact is right.