There is an increasing expectation in the academic sector for chemistry researchers to conduct risk assessment during experimental planning. However, information concerning laboratory scale chemical reactivity hazards can be di cult to parse despite ongoing e orts to compile from reported incidents. Laboratory procedures do not always directly ag possible incompatibilities among constituents or other process factors. In this paper, we present a pattern-based ontology for capturing multiple factors involved in laboratory procedures, including chemical properties, states, conditions, actions, and associated hazard classi cations.
Developing chemical safety risk assessment tools useful for the academic sector will necessitate tapping into digitally curated data in ways that are relevant to the decision-making processes of research chemists, safety professionals, institutional administration, and other stakeholders. For example, a researcher might be looking at two known chemicals in a proposed reaction scheme and want to know of any conditions that might trigger an adverse outcome, if there are any known procedures for minimizing the likelihood of these conditions, and how to mitigate potential harm if something untoward did occur. The relevant data and information may come from a diverse set of sources covering physical properties,3 synthesis protocols,4 and previously reviewed incidents,5 among other information.
Some of the most relevant information for analyzing risk appears in reports of incidents where safe control was exceeded, and the in uence of reactivity and process factors can be considered in retrospect. However, such reports are not the focus of normal research practice and tend to be exceedingly brief mentions found sporadically in letters to editors of journals,6 or as news items,7 or occasionally rephrased as caution statements in vetted procedures.8 Some of these reports have been collected into reference sources such as Bretherick's Handbook of Reactive Chemical Hazards, and the Pistoia Chemical Safety Library.9 Much of this content has been further compiled into an API-processable data stream within the PubChem database, dynamically presented in the Laboratory Chemical Safety Summaries format (LCSS)10 described by the US National Research Council (NRC) [3]. However, the meaning remains \locked" in unstructured text and not easily parsed for incorporation into digital information work ows.
The ability to make this information discoverable at the time of need will depend in part on more systematic description of these hazard scenarios. There are many factors at play in conducting a laboratory procedure that may contribute to the potential risk of a given situation. There is a body of research dedicated to analyzing the operations and conditions of large scale chemical processes in industrial settings, where these processes are well-de ned and carefully speci ed as part of the planning process [11].11 However, such analyses are rarely conducted for chemical procedures developed iteratively at the laboratory level as de ned by OSHA regulations in the United States. Analyzing procedures and coupling these with incident data can potentially bring to light incompatible combinations and problematic operations, as well as aid in planning for adjustments to experimental parameters. Domain terminology that describes key factors can enable the systematic analysis of relationships, such as combinations of chemicals, or substances under di erent conditions. Such approaches have been used for single analysis of M/SDS documents,12 and chemical procedures.13 Developing ontology patterns for chemical processes can more systematically represent potential intersections with hazardous situations [10].
Chemical information is predominantly organized by chemical entity, which is a limited perspective for discerning relationships among multiple process factors. The safety literature is no exception, focusing on hazard-related properties of individual chemicals or substances without reference to speci c experimental context or to the surrounding laboratory conditions. Scale, concentration, temperature, pressure, ow rate, and many other chemical, process, operator, and environmental factors have the potential to trigger \runaway" hazardous situations.14 A more complete risk assessment process, as described by the RAMP model, involves a holistic, laboratory level approach to managing risks beyond hazard identi cation [13]. Complementing the \object-based" index of speci c chemical entities with \process-based" modeling could help surface information and data buried in the published literature on how these chemicals are being used under various conditions and combinations, and the potential for subsequent unintentional interactions to arise [9]. 9 http://www.pistoiaalliance.org/projects/chemical-safety-library 10 https://pubchem.ncbi.nlm.nih.gov/lcss 11 www.acs.org/hazardassessment 12 www.ilpi.com/msds/ref/demystify 13 http://chemicaltagger.ch.cam.ac.uk 14 https://dchas.org/2017/04/05/information-flow-in-environmental-healthsafety
As such, we have begun the construction of a pattern ecosystem for capturing these chemical interactions and laboratory procedures. The foundational pattern is a chemical process pattern, which has been adapted from the State Transition pattern, which, in turn, is a generalization of the Semantic Trajectory pattern [7]. With the pattern, we hope to answer the following competency questions. 1. What substances appear in a particular action, together? 2. What substances are ever in the same container? 3. What temperatures or pressures are associated with these substances (conditions and/or changes)? 4. What apparatus or equipment is involved and associated with which substances (eg. glassware, stir-bars, glove-box) 5. What substances are co-located after some particular action? 2
In this section, we detail the Chemical Process Pattern. A graphical overview of the pattern can be seen in Figure 1. 2.1
The State Transition Pattern is a novel adaptation or modularization [5] of the Semantic Trajectory Pattern [7]. We provide a graphical representation of the pattern in Figure 1a.
The State Transition Pattern is a generalization of the Semantic Trajectory Pattern. The Semantic Trajectory deals with some Thing that moves through time and space which are captured as Fixes. In the State Transition Pattern, we have abstracted time and location to be Conditions of some State.
However, for our use case, we must further modularize the State Transition Pattern. At this time, the alignment is a set of subclass relations between the patterns, as follows.
Graphically, we see the results of these equivalences in Figure 1b. 2.2
Scoped Domain and Range. One of the primary goals of modelling with ontology design pattern is to lower the number of required ontological commitments required of an ontology engineer adopting the ontology. As such, we scope or guard many of the range and domain restrictions [6].
A v 8R.B
(
(b) We modularize the State Transition ODP to construct the Chemical Process pattern.
Axiom (
Structural Tautologies. These axioms are intended for human consumption; they do not add anything to the ontology. Essentially, these axioms, taking the below form, simply inform the reader of the intended use of a property [6].
A v
0R.B OPLa Annotations. The provided OWL le is annotated with the appropriate OPLa annoations [5]. We note, in particular, the classes marked as opla:ExternalClass: Action, Condition, and State. ChemicalActivity and EntityWithProvenance are dened later in the paper. The annotations were generated with the OPLa plugin for Protege [12].
Standard Disjointness. In the following sections, all classes which are not in direct or inferred subclass relationship are declared to be mutually disjoint. 2.3
Additionally, we provide graphical representations of the Stir Action and Heat Action subpatterns, as well as an expanded view of the Action Pattern in Figure 2. In the diagram, we use MethodTypes.txt and Apparatus.txt to denote that these values are individuals from a controlled vocabulary. An individual appearing the controlled vocabulary is an individual of type MethodType or Apparatus, for
> v 8withMethod.Method StirAction v =1withMethod.Method > v 8hasMethodType.MethodType
14. All StirActions are Actions. 15. The range of withMethod is strictly limited to Method. 16. A StirAction is completed with exactly one Method. 17. The range of hasMethodType is strictly limited to MethodType. 18. The domain of hasMethodType is strictly limited to Method.
HeatAction v =1untilTemperature.Temperature > v 8hasValue.Value
Temperature v =1hasValue.Value 19. All HeatActions are Actions. 20. A HeatAction has exactly one limiting Temperature. 21. The range of hasValue is strictly limited to Value. 22. A Temperature has exactly one Value. (14) (15) (16) (17) (18) (19) (20) (21) (22)
SimultaneousAction v Action 8hasSimultaneousAction.> v SimultaneousAction hasSimultaneousAction occursOver v occursOver hasSimultaneousAction involvesSubstance v involvesSubstance 23. All SimultaneousActions are Actions 24. The range of hasSimultaneousAction is strictly limited to Action. 25. A SimultaneousAction may not have another SimultaneousAction as a simultaneous action. 26. The domain of hasSimultaneousAction is strictly limited to SimultaneousAction. 27. The Actions that co-occur must, in fact, occur simultaneously. 28. Any Substance that is involved in a \subaction" is involved in the SimultaneousAction.
> v 8hasSimultaneousAction.Action > v 8hasSimultaneousAction.:SimultaneousAction (25) 2.4
ChemicalActivity v =1startsFrom.State ChemicalActivity v =1endsAt.State > v 8startsFrom.State > v 8endsAt.State 1. A ChemicalActivity always begins in some State and results in some State. 2. supra. 3. The range of startsFrom is strictly limited to States. 4. The range of endsAt is strictly limited to States. 2.5
> v 8hasAction.Action > v 8hasChemicalActivity.ChemicalActivity
ChemicalProcess v
1hasAction.Action
1hasChemicalActivity.ChemicalActivity
1hasState.State
(23)
(24)
(26)
(27)
(28)
(
1hasState.State > v 8hasState.State
1. A ChemicalSystem always has at least one State. 2. The range of hasState is strictly limited to State. 3. Any State is associated with exactly one Thing. 2.7
1. All Conditions must have provenance. In this use-case this is reasonable as every condition is measured by someone or some device. 2. The range of hasCondition is strictly limited to Conditions. 2.8
The EntityWithProvenance Pattern is extracted from the PROV-O ontology. At the pattern level, we do not want to make the ontological committment to a fullblown ontology. It su ces to align a sub-pattern to the core of PROV-O. Further discussion on the EntityWithProvenance pattern, as well as its speci cation (as below) in an OWL le may be found on the online portal.15 15 https://ontologydesignpatterns.org/wiki/Submissions:
EntityWithProvenance
(
1. The scoped range of wasDerivedFrom, scoped by EntityWithProvenance, is
EntityWithProvenance. 2. The scoped domain of attributedTo, scoped by Agent, is EntityWithProvenance. 3. The scoped range of attributedTo, scoped by EntityWithProvenance, is Agent. 4. The scoped domain of generatedBy, scoped by ProvenanceActivity, is EntityWithProvenance. 5. The scoped range of generatedBy, scoped by EntityWithProvenance, is ProvenanceActivity. 6. The scoped domain of used, scoped by EntityWithProvenance, is ProvenanceActivity 7. The scoped range of used, scoped by ProvenananceActivity, is EntityWith
Provenance.
8. The scoped domain of performedBy, scoped by Agent, is ProvenanceActivity.
9. The scoped range of performedBy, scoped by ProvenanceActivity, is Agent.
(
> v 8hasNextState.State 1hasNextState.State 1. The range of hasNextState is strictly limited to State. 2. A State will always follow at most one State. 3
The following incident report is extracted from [4, 1]. Formatting and language have been modi ed in order to make it clear exactly how the information was obtained. In the interest of brevity, we have used a simple incident report. However, even such a simple application of the pattern requires a high level of detail from the report. Thus, in our worked example, we aim to provide an illustration of the foundational concepts of our ontological ecosystem and note certain aspects will be addressed in future work. In the following, we use the cpp: namespace as an abbreviation for \Chemical Process Pattern" in the URI https://daselab.org/chemicalprocesspattern/.
The Incident Report. 5-ethyl-2-methyl-pyridine and 70% nitric acid were placed in a small auto-clave.
They were heated and stirred for 40 minutes.
The emergency vent was opened due to a sudden pressure rise. A violent explosion occurred 90 seconds later.
From the rst statement, we extract the following triples regarding the substances and apparatus. The placement of the chemicals will also constitute an Action subclass, as it is developed. cpp:sub1 rdf:type cpp:Substance
cpp:asText "5-ethyl-2-methyl-pyridine" . cpp:sub2 rdf:type cpp:Substance
cpp:asText "70% nitric acid" . cpp:ap1 rdf:type cpp:Apparatus
cpp:hasApparatusType "auto-clave" .
From the next sentence we extract the StirAction and HeatAction. In order to capture their simultaneity, we use the SimultaneousAction. cpp:te1 rdf:type cpp:TemporalExtent cpp:sa1 rdf:type cpp:StirAction cpp:ha1 rdf:type cpp:HeatAction cpp:sim1 rdf:type cpp:SimultaneousAction cpp:hasSimultaneousAction cpp:sa1 cpp:hasSimultaneousAction cpp:ha1 cpp:occursOver cpp:te1 . . . .
From the next sentence, we extract the apparatus and resulting state of the action. The Condition is provided an asText property for illustrative purposes. cpp:ap2 cpp:c1 cpp:s2 cpp:s1 rdf:type cpp:Apparatus cpp:hasApparatusType "fume hood" rdf:type cpp:Condition ewp:isAttributedTo cpp:ap2 cpp:asText "high pressure" rdf:type cpp:State rdf:type cpp:State cpp:hasNextState cpp:s2 . . . . . cpp:ca1 rdf:type cpp:startsFrom cpp:endsAt cpp:sim1 cpp:actsOn cpp:triggers cpp:ChemicalActivity cpp:s1 cpp:s2 cpp:s1 cpp:ca1
In the last step, we note that a hazardous state has been entered. However, the development of this part of the ontological ecosystem is still planned in future work. We note possible integration the Modi ed Hazardous Material Pattern [2] to help model this aspect. Finally, we may wrap it all together into the Chemical Process. cpp:cp1 rdf:type cpp:ChemicalProcess cpp:hasAction cpp:sa1 cpp:hasAction cpp:ha1 cpp:hasAction cpp:sim1 cpp:hasChemicalActivity cpp:ca1 cpp:hasState cpp:s1 cpp:hasState cpp:s2 . 4
In this paper, we have described a foundational pattern to building a ontology design pattern ecosystem for modelling chemical processes. The core pattern is based on the State Transition Pattern, which in turn, is adapted from the Semantic Trajectory Pattern. The intent of this pattern and the surrounding ecosystem is to provide chemists{and their students{ with a resource for analyzing experiments and potentially nding unforeseen interactions that can result in hazardous states, events, or situations.
A su ciently populated ontology of chemical processes can also be used as background knowledge for training a more sophisticated learning model or could be used to explain the decisions made by such a system (deep learning models and explainable AI, respectively).
In the future, we expect to integrate more closely with the large chemistry based datasets, such as PubChem and M/SDS. In addition, there are existing patterns that may be integrated to enhance the functionality of the core pattern and complete other pieces, such as QUDT16 for measurements and units, the Modi edHazardous Material Pattern [2] for modelling hazardous states, and the Material Transformation [8] for extending ChemicalActivity.
Acknowledgement. Cogan Shimizu acknowledges support by the Dayton Area Graduate Studies Institute (DAGSI). 16 https://qudt.org/