PaperlayerView exact revision on arXiv

Source-linked AI summary

SFE-O: An Ontology for Knowledge-Grounded AI in Supercritical CO$_2$ Biomass Extraction and Fractionation

Thomas Pallet, Mario Lezoche, Gianpaolo Leone, Daniele Pizzichini, Diego Torres, Herv{é} Panetto, M{é}lanie Glibusic

arXiv:2610.37123v1cs.AI

TL;DR

SFE process design and interpretation require heterogeneous, fragmented knowledge linking biomass properties, operating conditions, transport, and fractionation. SFE-O formalizes these connections and parameter dependencies in a modular ontology. A pilot-scale batch example was reasoner-checked and used to demonstrate competency-question querying, while numerical solving and AI evaluation remain outside the paper’s scope.

  • Problem

    SFE design and interpretation require heterogeneous knowledge about biomass, operating conditions, transport, and fractionation that is fragmented across documents, models, databases, and expert practice.

    Source: §I · pp. 1–2

  • Method

    SFE-O is a modular ontology that connects Biomass, Process, and Product concepts with process topology, parameter provenance, mass-transfer phenomena, CER/FER mechanisms, and computational dependencies.

    Source: §I · p. 2, §III · p. 3, B. Hydrodynamic and Mass-Transfer… · p. 7

  • Results

    A real pilot-scale batch was instantiated, classified as consistent by HermiT, and used to demonstrate that CQ1–CQ8 are expressible and queryable, not to benchmark performance.

    Source: §VII · pp. 7–8, §VIII · p. 8

  • Takeaways & Limitations

    SFE-O provides a machine-actionable semantic layer for querying, consistency checking, and traceable mechanistic explanation, with ontology-grounded AI proposed as a future direction.

    Source: §I · p. 2, §VI · p. 7, §VIII · p. 8

    Numerical dependencies are not solved within OWL, external services remain integration points, and effective diffusivity requires biomass-specific assumptions.

    Source: §VII · p. 8

Abstract

from arXiv · show

Supercritical CO$_2$ extraction (CO$_2$-SFE) offers a sustainable route for recovering high-value compounds from biomass while limiting the use of conventional organic solvents, but its design and interpretation require heterogeneous knowledge spanning biomass properties, operating conditions, mass transfer, and product fractionation, typically fragmented across technical reports, engineering models, external databases, and expert practice. This paper introduces SFE-O, a modular Web Ontology Language (OWL 2) ontology for supercritical CO$_2$ biomass extraction, structured around interconnected Biomass, Process, and Product dimensions and extended with explicit representations of process topology, parameter provenance, external and internal mass-transfer phenomena, and the Constant Extraction Rate (CER) / Falling Extraction Rate (FER) mechanisms. Computational dependencies among process quantities are represented explicitly, without embedding numerical solving in the ontology, so that input, measured, computed, externally retrieved, and fitted parameters remain traceable. SFE-O reuses established semantic resources (ChEBI, QUDT, OWL-Time, PROV-O) and is evaluated through a worked example that instantiates a real pilot-scale extraction batch, a reasoner-based consistency check, and a competency-question SPARQL query. The resulting knowledge model provides a machine-actionable foundation for querying, consistency checking, and mechanistic explanation of SFE processes, and establishes a knowledge layer that could support future ontology-grounded AI reasoning, a direction this paper motivates but does not evaluate.

I. INTRODUCTION

CO2-SFE can recover valuable compounds under comparatively mild conditions, but designing and interpreting it requires fragmented knowledge linking biomass, process conditions, transport, and fractionation. SFE-O formalizes those links as a traceable semantic layer for SFE knowledge.

I. INTRODUCTION

  • SFE behaviour depends on interrelated biomass properties, operating conditions, transport mechanisms, and separator configurations, with pilot-scale fractionation using sequential stages at different pressure and temperature conditions.The knowledge is distributed across measurements, calculations, external databases, models, and expert practice.Source: §I · pp. 1–2
  • Existing knowledge is fragmented, and general process ontologies do not connect biomass characteristics, SFE mechanisms, fractionation, derived quantities, and their provenance.The Broken-and-Intact Cell approach links disrupted and intact biomass structures to fast, externally controlled extraction and subsequent internal-diffusion effects,.Source: §I · p. 2
  • SFE-O organizes extraction knowledge into interconnected Biomass, Process, and Product dimensions and represents topology, temporal phases, transport, computational dependencies, and parameter provenance.It distinguishes input, measured, computed, externally retrieved, and fitted values, and formalizes links between biomass disruption and CER/FER mechanisms.Source: §I · p. 2
  • SFE-O is a machine-actionable semantic layer for querying, consistency checking, traceable mechanistic explanation, and interoperability, rather than a replacement for numerical SFE models.The work focuses on knowledge formalization and ontology validation and does not evaluate experimental datasets.Source: §I · p. 2

II. RELATED WORK

  • Prior work spans mechanistic extraction modelling, semantic process and biomass models, and reusable chemical and metrological ontologies, but their integration for SFE is not addressed.The paper positions SFE-O at the intersection of these research areas.Source: §II · p. 2

A. Process and Biomass Ontologies

Process and biomass ontologies provide reusable engineering and biorefinery concepts, while chemical, metrological, temporal, and provenance resources supply complementary semantics. SFE-O builds on these resources to represent SFE-specific interactions.

A. Process and Biomass Ontologies

  • OntoCAPE models broad process-engineering concepts, while BiOnto and related biomass models address feedstocks, composition, processing pathways, and biorefinery streams.These approaches do not specifically represent the SFE mechanisms and interactions among morphology, topology, transport, kinetics, and fractionation.Source: A. Process and Biomass Ontologies · p. 2

B. Reusable Semantic Resources

  • SFE-O can reuse ChEBI, FoodOn, QUDT, OWL-Time, and PROV-O to represent chemical entities, food-related products, quantities and units, temporal structure, and parameter provenance.These resources complement rather than replace SFE-specific knowledge about how those dimensions interact in an extraction process.Source: B. Reusable Semantic Resources · p. 3, A. Process and Biomass Ontologies · p. 3

C. Research Gap and Positioning of SFE-O

Existing ontologies cover process systems, biomass pathways, and individual semantic dimensions, but do not integrate them into a machine-actionable representation of SFE mechanisms. SFE-O was developed through iterative knowledge elicitation to connect these domain dimensions and their dependencies.

C. Research Gap and Positioning of SFE-O

  • Existing process, biomass, chemical, and metrological ontologies do not integrate their concepts around SFE mechanisms, and the authors know of no ontology combining BIC kinetics with explicit parameter provenance.SFE-O specializes the extraction domain while reusing established semantics for traceable reasoning and future knowledge-grounded AI applications.Source: C. Research Gap and Positioning of SFE-O · p. 3

III. KNOWLEDGE ELICITATION AND ONTOLOGY REQUIREMENTS

  • SFE-O construction adapted an iterative ontology-development cycle and competency-question method to identify concepts, relations, constraints, and computational dependencies rather than reproduce experimental datasets.Domain knowledge was progressively extracted, structured, reconciled, and formalized.Source: §III · p. 3

A. Knowledge Sources and Elicitation Process

  • Six heterogeneous knowledge artefacts supplied complementary terminology, pilot-scale topology, parameter correspondences, computational dependencies, hierarchies, parameter origins, and mechanism definitions.The sources included technical reports, a parameter reference, a computational spreadsheet, conceptual diagrams, and an expert-authored mechanism model; no one source contained all required knowledge.Source: A. Knowledge Sources and Elicitation… · pp. 3–4
  • The elicited model links Biomass, Process, and Product perspectives through causal and computational dependencies rather than treating them as a sequential taxonomy.They cover feedstock composition and morphology, extraction equipment and conditions, and products and fractions with quantities such as yield, recovery, and concentration.Source: A. Knowledge Sources and Elicitation… · p. 4

B. Ontology Requirements

SFE-O’s requirements connect its three domain perspectives while preserving reusable semantics, contextual quantities, parameter origins, mechanistic dependencies, and product-data distinctions.

B. Ontology Requirements

  • R1–R2 organize the ontology into interconnected Biomass, Process, and Product modules and reuse ChEBI, QUDT, OWL-Time, and PROV-O instead of re-formalizing their semantics.R4 and R7 keep chemical roles contextual and separation topology generic, while R8 distinguishes missing values, below-detection measurements, and measured zeroes.Source: B. Ontology Requirements · p. 4
  • R3 models concentrations as contextual quantities linked to a chemical entity and material, with numerical value, unit, and any needed reference basis.This avoids adding schema properties for each new compound.Source: B. Ontology Requirements · p. 4
  • R5 records whether parameters are set points, measured, calculated, externally retrieved, or fitted, supporting traceability of how results were obtained.Examples include pressure, biomass composition, Reynolds or Sherwood numbers, CO2 properties, broken-cell fraction, and effective diffusivity.Source: B. Ontology Requirements · pp. 4–5
  • R6 represents computational dependencies such as Reynolds and Schmidt numbers contributing to Sherwood number and the external mass-transfer coefficient, while leaving numerical evaluation to external software or services.For CER/FER, broken-cell fraction g is associated with faster external control, while intact fraction (1 −g) contributes internal resistance and diffusion-controlled extraction.Source: B. Ontology Requirements · p. 5

C. Competency Questions

The competency questions define which biomass, process, transport, product, provenance, and constraint information SFE-O should make accessible. They guide the ontology’s architecture, reuse decisions, properties, and constraint representations.

C. Competency Questions

  • The questions cover relevant biomass characteristics and chemical entities, plus the geometric and operating quantities of extraction batches and stages.Source: C. Competency Questions · p. 5
  • They ask which quantities support deriving Re, Sc, Sh, and kf, and whether each was supplied, measured, calculated, externally retrieved, or fitted.Source: C. Competency Questions · p. 5
  • They ask how the broken-cell fraction relates to CER- and FER-associated target compound portions.Source: C. Competency Questions · p. 5
  • They cover which separation stage generates each product fraction and under what pressure and temperature, as well as fraction chemistry and functional roles.Source: C. Competency Questions · p. 5
  • A final question asks whether represented quantities and relationships satisfy encoded domain constraints, and the set defines what knowledge must be explicitly accessible.Source: C. Competency Questions · p. 5

IV. SFE-O ARCHITECTURE AND SEMANTIC REUSE

SFE-O organizes extraction knowledge into seven local OWL 2 modules and an aggregate ontology, with connections to external semantic resources. Its modular architecture represents biomass, process, transport, product, provenance, and application knowledge.

IV. SFE-O ARCHITECTURE AND SEMANTIC REUSE

  • The ontology comprises Core, Biomass, Process, Transport, Product, Provenance, and Application modules, plus an aggregate ontology importing the complete model.Fig. 1 summarizes the architecture and its connections with external resources.Source: §IV · p. 5, Figure 1 · p. 6

A. Modular Architecture

  • The biomass module represents batches, pretreatment, morphology, physical characteristics, chemical composition, and broken versus intact structural units.Source: A. Modular Architecture · p. 5
  • The process module distinguishes an ExtractionBatch, the complete operation on a biomass load, from an ExtractionCycle, one CO2 circulation through functional stages.This distinction supports single-pass laboratory systems and closed-loop pilot or industrial configurations.Source: A. Modular Architecture · p. 5
  • Extraction cycles can contain ordered SeparationStage individuals linked to equipment, pressure, temperature, and generated ProductFraction, without limiting the model to a three-separator plant.Source: A. Modular Architecture · p. 5
  • Transport represents diffusivity, Reynolds, Schmidt, and Sherwood numbers, kf, and CER/FER mechanisms; product and provenance modules represent fractions and outcomes, and parameter origins.The application module holds application-level SFE knowledge without representing experimental observations.Source: A. Modular Architecture · pp. 5–6

B. Semantic Modelling Patterns

SFE-O uses three modelling patterns to represent chemical composition, contextual process roles, and computational dependencies. These patterns preserve semantic context and derivation relationships without requiring OWL to execute numerical models.

B. Semantic Modelling Patterns

  • Chemical composition is represented as a reified relation linking a material, chemical entity, and quantity, with an optional reference basis such as dry or wet mass.New compounds can be introduced without modifying the ontology schema.Source: B. Semantic Modelling Patterns · p. 6
  • Chemical entities are not intrinsically solutes; they can play a SoluteRole in an extraction context, as solvent and co-solvent roles also depend on process configuration and conditions.Source: B. Semantic Modelling Patterns · p. 6
  • Computational dependencies link calculations to required quantities and outputs without requiring OWL to execute numerical models.For example, Sherwood number depends on Reynolds and Schmidt numbers, while kf depends on Sherwood number, solute diffusivity, and particle size; engineering software performs numerical evaluation while the ontology preserves dependencies and provenance.Source: B. Semantic Modelling Patterns · p. 6

C. Reuse of External Semantic Resources

SFE-O selectively reuses or aligns established semantic resources rather than recreating their knowledge locally. This keeps the ontology focused on SFE-specific knowledge while supporting interoperability.

C. Reuse of External Semantic Resources

  • SFE-O delegates quantities and units to QUDT, temporal concepts to OWL-Time, provenance to PROV-O, and aligns chemical entities with ChEBI.NCBI Taxonomy, FoodOn, and LIPID MAPS serve as interoperability or extension points rather than full imports.Source: C. Reuse of External Semantic Resources · p. 6
  • Delegating chemical identity, units, time, and provenance to established resources keeps SFE-O focused on knowledge specific to supercritical CO2 extraction.Source: C. Reuse of External Semantic Resources · p. 6

V. MECHANISTIC KNOWLEDGE AND CONSTRAINT FORMALIZATION

SFE-O explicitly connects biomass structure and operating conditions to mass-transfer mechanisms and product formation. It represents CER/FER behavior and computational dependencies while preserving parameter provenance, including biomass-specific dependencies without imposing a universal effective-diffusivity equation.

V. MECHANISTIC KNOWLEDGE AND CONSTRAINT FORMALIZATION

  • SFE-O explicitly represents mechanistic relationships connecting biomass structure, operating conditions, mass transfer, and product formation.This links the domain’s major dimensions rather than treating process quantities as isolated parameters.Source: §V · p. 7

A. CER/FER Representation

  • The BIC interpretation distinguishes an accessible solute fraction in disrupted structures from a less accessible fraction retained in intact structures.The representation uses loaded biomass mass, initial compound concentration, and broken-cell fraction to describe the amounts.Source: A. CER/FER Representation · p. 7
  • The broken-cell fraction is associated with CER, predominantly governed by external mass transfer, while intact cells contribute to FER, increasingly controlled by internal diffusion.The two extraction regions correspond to distinct transport mechanisms.Source: A. CER/FER Representation · p. 7
  • Extraction efficiency and concentration can be linked to the mechanistic model through total extracted mass, initial compound amount, and total extracted material.The indicators are EFF = MTOT/C0(i) and CONC = MTOT/EA.Source: A. CER/FER Representation · p. 7

B. Hydrodynamic and Mass-Transfer Dependencies

  • Pressure and temperature determine CO2 density and viscosity, while biomass morphology and molecular identity contribute particle size and compound diffusivity.Together with flow and bed characteristics, these quantities determine Reynolds and Schmidt numbers.Source: B. Hydrodynamic and Mass-Transfer… · p. 7
  • SFE-O records each derivation as a CalculationActivity and distinguishes parameter fitting from external lookups.Effective diffusivity depends on biomass-specific porosity, tortuosity, and microstructure, but the elicited sources provide no universal equation, so the dependency is recorded without imposing one.Source: B. Hydrodynamic and Mass-Transfer… · p. 7

C. Constraints

The constraints separate OWL logical reasoning from numerical constraint execution, while the AI perspective assigns semantic grounding to SFE-O and calculation to numerical services. Evaluation demonstrates the ontology’s consistency and queryability on a real pilot-scale batch, while identifying limits in numerical execution and experimental scope.

C. Constraints

  • OWL axioms express logical conditions, while arithmetic bounds, parameter ranges, and derived-quantity relations remain machine-readable expressions not evaluated by the reasoner.The ontology also distinguishes not analysed, below detection limit, and measured zero as separate states.Source: C. Constraints · p. 7

VI. KNOWLEDGE-GROUNDED AI PERSPECTIVE

  • The proposed AI architecture assigns admissible concepts, relations, and dependency chains to SFE-O, calculations to numerical services, and interaction to an LLM.No LLM system is built or evaluated in this paper.Source: §VI · p. 7

VII. EVALUATION AND DISCUSSION

  • Evaluation checks OWL parsing and imports, assesses requirements and competency questions, and reviews representation of process, transport, separation, and parameter provenance.SFE-O is evaluated as a knowledge representation rather than as a predictive system.Source: §VII · p. 7
  • The instantiated pilot-scale ontology was classified as consistent by the HermiT DL reasoner, including its qualified-cardinality restriction.The example uses Dunaliella salina operating data and a three-separator cascade.Source: §VII · p. 7
  • A worked pilot-scale batch produced an exact source-value match in a SPARQL competency query, demonstrating queryability on real data rather than benchmark performance.The example uses a three-separator cascade and asks which stage generated a fraction and under what pressure and temperature.Source: §VII · pp. 7–8
  • Numerical dependencies are not solved within OWL, constraints are not yet executable SHACL shapes, and external-property, diffusivity, and fitting services remain integration points.Effective diffusivity also requires biomass-specific assumptions; several source-data inconsistencies remain flagged for expert validation.Source: §VII · p. 8
  • The paper excludes broader experimental observations beyond its single worked example and calls for evaluation on additional batches and competency questions.Future work should also test executable constraints and whether ontology-grounded interaction improves traceability, consistency, explanation, and process-space exploration.Source: §VII · p. 8

VIII. CONCLUSION

SFE-O formalizes interconnected SFE knowledge, including BIC/CER–FER mechanisms and computational dependencies, and demonstrates the ontology on an instantiated extraction batch.

VIII. CONCLUSION

  • SFE-O explicitly formalizes BIC/CER–FER mechanisms and computational dependencies while connecting biomass, process topology, transport, separation, products, and provenance.It reuses ChEBI, QUDT, OWL-Time, and PROV-O.Source: §VIII · p. 8
  • An instantiated extraction batch demonstrates SFE-O, with a HermiT-checked full OWL 2 DL consistency result and an executed competency-question SPARQL query.Source: §VIII · p. 8
  • Future work will add executable constraint validation, numerical and physicochemical services, experimental instantiation, and evaluation of ontology-grounded AI for process interpretation and optimisation.Source: §VIII · p. 8
Loading 2610.37123v1…