Skip to content

Design by analogy and computational creativity

Part of Inverse Innovation with the Encyclopedia of Abstractions · Design by analogy and computational creativity · Last revised August 2026

Review date: 2026-08-03
Coverage: sources discoverable through 2026-08-03
Claim status: cross-disciplinary scoping review; nearest-system and claim-boundary memo, not a priority claim

Executive conclusion

The literature contains strong precedents for nearly every individual operation in the Encyclopedia of Abstractions (EoA) pipeline under review. It also contains several important precedents for the less common direction from a known solution, technology, or biological strategy toward new problems or applications. The strongest contrary case found is AskNatureGPT: it begins with a biological solution, automatically identifies an engineering problem/application, transfers the biological analogy, generates a natural-language design concept, and applies two learned relevance evaluators. Patent-based technology-opportunity-discovery (TOD) systems likewise start from existing technologies or products and systematically infer possible applications at much larger corpus scale. Human technology–market-linking research has followed existing technologies through the discovery, development, and validation of need–solution pairs.

The evidence therefore does not support describing solution-to-problem transfer, analogical problem finding, structured mechanism representation, automated concept generation, or large-scale opportunity search as historically unprecedented. It does support a narrower provisional gap: within this scoped search, I did not find a demonstrated system that combines all of the following in one reproducible workflow:

  1. domain-general, explicitly curated solution archetypes and mechanisms;
  2. a predeclared archetype-by-target-domain matrix rather than an ad hoc query or technology neighborhood;
  3. multiple operational proposals per pairing;
  4. external prior-art checking of the generated proposal (not merely use of patents as an inspiration corpus);
  5. independent or adversarial criticism separated from generation;
  6. practicality, cost, adopter, risk, and impact scrutiny;
  7. a falsifiable empirical next step;
  8. iterative repair with preserved rejected and terminal outcomes; and
  9. reproducible raw artifacts, provenance, and resource accounting suitable for a synthetic training curriculum.

That is an integration and experimental-design distinction, not evidence of historical priority. AskNatureGPT, function-based TOD, analogical scientific search, Idea-Inspire/DANE, InnoGPS, AutoTRIZ, and deliberate need–solution search should be treated as substantive baselines, not background decoration.

Taxonomy of the field

Direction of work

  • Problem → source solution → adapted solution. This is the dominant design-by-analogy pattern: formulate a target need, retrieve functionally or causally similar cases, map source relations, and adapt a candidate. Idea-Inspire, DANE, AskNature, Analogy Finder, SOLVENT/scientific analogy search, and AutoTRIZ mainly operate here.
  • Solution/technology → problem/application. Solution-driven biologically inspired design, outward technology transfer, and capability-based TOD begin with a biological strategy, patent, technology, or product and seek needs it could address. AskNatureGPT, Yoon et al.'s function-based TOD framework, Qiao et al.'s technology–function link prediction, and the organizational search studied by Reinsberger et al. are central precedents.
  • Problem–solution co-evolution. Design cognition often alternates between changing a problem formulation and changing candidate solutions. Helms and Goel's “analogical problem evolution” is especially relevant because existing source solutions can change the target problem even without immediately producing a new solution.
  • Concept/representation → new artifact. Conceptual blending, bisociation, and computational creativity systems combine or transform represented concepts. They can generate unfamiliar artifacts or concepts, but usually lack a target-domain problem-discovery and deployment-validation pipeline.

Representation families

  • Relational analogy. Gentner's structure-mapping theory privileges systematic relational correspondences over object attributes; Holyoak and Thagard's ACME adds interacting structural, semantic, and pragmatic constraints. These theories explain why far analogy cannot be reduced to lexical similarity.
  • Function and functional basis. A standardized verb–object/function–flow vocabulary makes artifacts searchable independently of form. It supports patent retrieval, biological-to-engineering translation, and repository reuse, but function alone is a weaker representation than a causal mechanism.
  • FBS/SBF. Gero's Function–Behaviour–Structure (FBS) separates intended purpose, behavior derivable from structure, and the components/relations constituting structure. DANE's Structure–Behavior–Function (SBF) models capture causal state transitions in biological systems and index systems by function.
  • SAPPhIRE. State change, Action, Part, Phenomenon, Input, oRgan, and Effect provide a multi-level causal account. Idea-Inspire uses this representation for structured retrieval and explanation; later work has used it to structure analogies and assess problem novelty against a reference problem database.
  • TRIZ representations. Technical/physical contradictions, inventive principles, substance–field models, and technological-evolution patterns encode generalized problem–solution regularities derived from patent analysis. The classic workflow is specific problem → generalized contradiction → generalized solution principle → specific solution.
  • Purpose–mechanism schemas and embeddings. Analogy-mining systems learn lightweight “what is it for?” and “how does it work?” representations from products or papers. They trade representational depth for million-document scale.
  • Patent opportunity networks. Products, technologies, functions, IPC classes, subject–action–object (SAO) triples, citations, or semantic phrases become nodes and edges; missing or distant links are interpreted as possible opportunities.
  • Conceptual spaces and blends. Bisociation and conceptual blending combine previously separate frames; computational implementations such as Divago and COINVENT formalize mapping, generalization, blend construction, constraint checking, or argumentation. They are primarily concept generators rather than problem–solution validation systems.

Chronological and technical synthesis

Generalized solution knowledge before modern analogy engines

TRIZ is the clearest early example of systematically abstracting reusable solution patterns from inventions. Altshuller and colleagues organized recurring patent-derived strategies into contradiction models, inventive principles, ARIZ, substance–field analysis, and laws/patterns of technical-system evolution. Its importance here is architectural: it explicitly separates a concrete problem from a generalized problem and a generalized solution. However, classical TRIZ is normally problem initiated. Its systematicity is over contradiction categories and principles, not a cross-product of solution archetypes and unrelated target domains. Its patent foundation is also not equivalent to running a contemporary prior-art search on each generated proposal.

Koestler's 1964 account of bisociation framed creativity as bringing previously separate matrices of thought together. Fauconnier and Turner's later conceptual-blending theory formalized interacting input spaces, a generic space, cross-space mappings, and an emergent blend. These are foundational accounts of cross-domain combination, but neither by itself supplies an innovation-governance pipeline.

Analogy, cases, and explicit design representations

Gentner's 1983 structure-mapping theory established the central distinction between deep relational similarity and shared attributes. Holyoak and Thagard's 1989 ACME program demonstrated computational mapping under structural, semantic, and pragmatic constraints. Case-based and derivational reasoning added retrieval of past episodes, replay of prior solution paths, adaptation, and learning from success or failure. Kolodner and Wills explicitly connected case-based reasoning to creative design, including evaluation-driven refinement and reformulation of specifications.

Gero's 1990 design-prototype schema made function, expected and actual behavior, structure, context, and relational/qualitative/computational knowledge explicit. The later situated FBS framework modeled designing as transformations and reformulations among function, behavior, structure, and descriptions in a changing context. FBS is an ontology and process account, not intrinsically an analogy engine; its contribution is to prevent “same function” from standing in for “same mechanism.” Critical analysis has also questioned the boundary between intentional and structural descriptions and the empirical status of variants of the model.

The engineering Functional Basis standardized form-independent function–flow descriptions. WordTree expanded design-problem language to find distant analogies without a specialized repository. Function-based patent search then automated extraction and vector matching. Murphy et al. showed retrieval ranging from literal to far-field patent analogies; subsequent analogy-mining work learned purpose and mechanism representations from 8,500 crowdsourced product descriptions and experimentally improved creative ideation relative to surface retrieval.

Bio-inspired and biomimetic design systems

Bio-inspired design supplies unusually clear evidence for both directions of transfer. Problem-driven work starts with an engineering need and searches biology. Solution-driven work starts with a biological system and seeks engineering problems or applications. Empirical work by Helms and Goel showed that analogues can alter the problem formulation itself, not merely donate a solution.

Key systems differ in representational depth:

  • AskNature is a public, function-indexed repository of biological strategy pages, contextual search, and bio-inspired products. It abstracts transferable biological information but generally leaves mapping and engineering concept development to the designer.
  • DANE captures biological systems in SBF models and supports function-indexed retrieval and causal understanding. Its strength is explanatory depth; its limits include costly knowledge engineering, a curated biology scope, and human-led adaptation.
  • Idea-Inspire 3.0/4.0 uses SAPPhIRE-based multi-system representations, structured search, and multimodal explanations. Controlled studies test understanding and application of biological concepts to engineering problems. It is among the closest mechanism-explicit analogical design environments, but concept generation and evaluation remain substantially human and problem led.
  • Function-based biologically inspired design translates biological systems through a Functional Basis and repositories, with concept-generation software able to associate biological and engineering components. This demonstrates that functional abstraction can enable solution-driven concept generation, though it does not supply broad problem discovery, prior-art checking, or governance.
  • BioTRIZ maps biological and technological strategies through TRIZ dimensions and showed that biology and technology occupy different solution tendencies. It is a principled bridge across domains, but still a prescriptive ideation method rather than an end-to-end screening and validation system.
  • PeTaL/BIDARA is an open NASA project using AI and biomimicry knowledge to help users solve stated engineering challenges. It is relevant for artifact openness and guided transfer but is principally problem driven.

Patent analogy and technology-opportunity discovery

Patent systems split into two families that must not be conflated.

Analogy retrieval finds prior solutions to a supplied need. Function-vector search, Analogy Finder, learned purpose–mechanism schemas, and focused abstraction can retrieve distant but relevant products or patents. Gilon et al.'s focus-abstracted queries returned results as relevant as surface baselines while being more domain distant. These systems demonstrate scalable retrieval and some structural abstraction, not automated adaptation or proposal adjudication.

Technology-opportunity discovery can run in the outward direction. Yoon et al. built a function-based knowledge base from 223,603 patents and defined four paths from existing technologies/products to modifiable technologies, producible products, related products, or technologies applicable to a product. InnoGPS maps more than five million patent records into a technology space, supports positioning, neighborhood exploration, path finding, and near/far-domain inspiration, and has been used to conceive applications and retrieve domain-specific patent concepts. Qiao et al. extract technology–function pairs with SAO parsing, build a bilayer network, map explicit applications, and use link prediction for implicit applications. These are direct precedents for systematic solution/technology-to-application search.

Their output, however, is generally an opportunity, product/technology relation, ranked application field, or inspiration—not a complete domain-grounded intervention with stakeholder analysis, costs, falsifiable test, independent review, repair history, and a terminal disposition. Using patents to construct a search space also does not establish that the generated opportunity has passed legal novelty or freedom-to-operate review.

Computational blending and computational creativity

Divago maps two represented domains and generates a blend under pragmatic and optimality constraints. COINVENT broadened this agenda into formal concept invention across ontologies, logic, mathematics, and music, including work on argumentation to evaluate blends. WHIM generated and evaluated fictional “what-if” ideas. These systems establish that structured cross-domain combination, automatic generation, and machine evaluation predate LLM systems.

Their relationship to EoA is adjacent rather than end-to-end. They generate concepts or artifacts in a chosen creative space, but generally do not discover target-domain problems from a curated solution archetype, establish target need and adopters, conduct external prior-art search, develop an operational intervention and empirical test, or retain a resource-accounted critique-and-repair dossier. Their evaluators also usually encode domain-internal aesthetic, coherence, novelty, or value criteria rather than independent deployment criticism.

Scalable scientific analogy and LLM-era systems

Hope et al. learned purpose–mechanism “problem schemas” from product descriptions and showed that lightweight structure can improve analogical retrieval and downstream ideation. Kang et al. extended the approach to approximately 1.7 million scientific papers across computer science, engineering, biomedicine, and business/engineering. Their search engine extracts purpose and mechanism spans, retrieves papers that address related purposes with different mechanisms, supports human filtering, and was later automated. Scientists produced more creative adaptations from analogy results than direct applications, and the authors released purpose/mechanism embeddings. This is the strongest precedent for scalable, artifact-backed retrieval and human adaptation in science; it begins with a scientist's problem and stops short of autonomous proposal validation.

AutoTRIZ uses multiple LLM modules plus a fixed TRIZ knowledge base to turn a user problem statement into an interpretable report with multiple solutions. It automates a much larger portion of the problem-to-solution path than classic tools and reports consistency experiments plus textbook and battery-thermal-management cases. It remains problem initiated, and its evaluation does not constitute external prior-art checking or independent empirical validation.

AskNatureGPT is the closest computational contrary case. It fine-tunes GPT-3.5 on 305 successful AskNature/AskNatureNet cases. Given a biological source, benefits, and biological model, the generator identifies an engineering application/problem and generates an innovation description. Two separately fine-tuned classifiers assess whether the output remains related to the biological benefit and model. The study reserved 25 cases for testing, generated ten outputs for each, used ablations and semantic metrics, conducted human novelty/feasibility assessment, and presented three design cases. It therefore demonstrates solution-to-problem generation, analogical transfer, complete natural-language concepts, multiplicity, machine screening, held-out evaluation, and a reusable supervised dataset.

Its limits matter for the claim boundary. The source space is biological rather than domain-general; the 305 examples are successful BID cases, not a factorial archetype-by-domain corpus; the two evaluators test source–output relevance and are derived from the same base-model/data family, not independent adversarial critics; negative examples are randomized mismatches rather than substantive failed proposals; the quantitative evaluation supplies the known application title to constrain generation; and the paper does not report external prior-art review, adopter/cost analysis, built-prototype testing, iterative repair trajectories, or resource accounting.

Finally, Reinsberger et al. provide a strong non-computational contrary case. Their 2026 Research Policy study follows four organizational technology–market-linking projects through deliberate exploratory search for novel applications of existing technologies, using 306 expert interviews, 89 innovation proposals, and longitudinal evidence from 18 search agents. Needs and solutions co-evolve into need–solution pairs through domain crossing, evaluation, and learning. This is closer to real deployment and adopter evidence than most computational work, but it is an organizational process study rather than a reproducible archetype-by-domain generation system.

Nearest-system matrix

Legend: Y demonstrated; P partial, human-mediated, narrow, or indirect; N not demonstrated in the reviewed source; NR not reported. “Patent corpus” is not scored as external prior-art checking unless the generated output itself is searched and adjudicated.

Protocol dimension AskNatureGPT Function-based TOD (Yoon/Qiao) Scientific analogy engine Idea-Inspire / DANE InnoGPS AutoTRIZ Divago / COINVENT Deliberate need–solution search
Explicit reusable solution archetype P N P P N Y P N
Explicit mechanism or structural-role mapping P P Y Y P P Y P
Semantically distant transfer Y P Y Y Y P Y Y
Solution-to-problem generation Y Y P P Y N N Y
Complete domain-specific proposal Y P P P P Y N Y
Multiple proposals per pairing Y Y Y P Y Y Y Y
Predeclared archetype-by-domain matrix N N N N N N N N
Independent/adversarial criticism N N P P N N P P
External prior-art search on outputs N N N N N N N NR
Practicality, cost, adopter, impact assessment P P P P P P N Y
Falsifiable empirical next step in each proposal N N N N N N N P
Iterative repair with preserved trajectory N N N P N N P P
Negative examples and terminal classifications P P P N N N P P
Reproducible raw artifacts and resource accounting P NR Y P NR P P NR
Explicit synthetic training curriculum Y N P N N N N N

What “nearest” means here

  • Closest directional and generative precedent: AskNatureGPT.
  • Closest solution-to-application system at patent scale: Yoon et al.'s function-based TOD, with Qiao et al.'s link-prediction extension.
  • Closest retrieval system at scientific-corpus scale: Kang et al.'s analogical search engine.
  • Closest deep causal representation and pedagogical transfer environment: Idea-Inspire and DANE/SR.BID.
  • Closest global technology-space opportunity navigator: InnoGPS.
  • Closest automated generalized-principle-to-specific-solution pipeline: AutoTRIZ.
  • Closest formal concept-combination tradition: Divago/COINVENT.
  • Closest real organizational solution-to-need development process: Reinsberger et al.

No one of these labels means “overall predecessor” or “inferior system.” Each optimizes a different part of the task.

Strongest contrary cases

1. AskNatureGPT

This source directly contradicts any broad claim that computers have not generated problems and domain-specific concepts from source solutions. Its explicit sequence—biological solution → engineering application/problem → innovation description—and its learned evaluators make it the mandatory nearest baseline. The correct distinction is the breadth and governance of the pipeline, not directionality alone.

2. Function-based technology-opportunity discovery

Yoon et al. and Qiao et al. contradict claims that systematic, large-data solution-to-application search is absent. They formalize functions across hundreds of thousands of patents and infer explicit or missing technology–function links. EoA must distinguish curated domain-general archetypes and mechanism-preserving proposals from capability- and co-occurrence-based opportunity graphs, then test whether that distinction yields better opportunities.

3. Analogical problem evolution and SR.BID

Helms and Goel contradict a strict separation between problem discovery and solution transfer. Existing biological solutions can change or originate a problem formulation; SR.BID explicitly represents problem specifications and supports comparison of problems and systems. EoA should present solution-to-problem work as a formalized and scaled continuation of known bidirectional design cognition, not a new direction of reasoning.

4. Deliberate exploratory search for need–solution pairs

Reinsberger et al. show that organizations can systematically search for applications of existing technologies and iteratively validate emerging need–solution pairs. This is a stronger real-world comparator for adopter and viability work than most AI ideation systems. EoA's advantage, if demonstrated, would concern reproducible coverage, explicit structure, preserved alternatives, and experimental controls—not the existence of technology-push search.

5. SOLVENT and the 1.7-million-paper analogy engine

These systems contradict claims that structural analogical retrieval cannot scale or that distant scientific mechanisms cannot be retrieved and adapted empirically. Their released representations and user studies are important baselines for retrieval quality, domain distance, and downstream creative adaptation. Their weakness relative to the protocol is downstream proposal scrutiny, not retrieval.

Implications for EoA claims and evaluation

Language supported by this review

  • “The pipeline combines established methods from analogical design, technology-opportunity discovery, computational creativity, and structured evaluation in a predeclared cross-domain experiment.”
  • “Within the scoped search, we found close partial precedents but no single demonstrated system matching the full configured workflow.”
  • “The contribution tested here is integration, systematic coverage, provenance, and scrutiny—not the invention of analogy, solution-driven design, or computational problem finding.”
  • “The approach uses curated solution archetypes and mechanisms as experimental units, whereas many prior systems use cases, functions, contradictions, patents, biological strategies, or learned purpose–mechanism embeddings.”

Avoid “first,” “unique,” “unprecedented,” “no prior system,” and “inverse innovation has not been attempted.” Also define the term inverse innovation locally: “reverse innovation” already names a different international-management phenomenon, while “inverse design” commonly denotes optimization from desired behavior to structure.

Required baselines

  1. AskNatureGPT-style solution-driven generation: source solution/benefits → candidate application and concept, with a relevance evaluator.
  2. Function-link TOD baseline: extract functions and rank target applications from existing source capabilities without curated archetypes.
  3. Purpose–mechanism retrieval baseline: retrieve distant cases from a broad corpus, followed by the same proposal generator.
  4. AutoTRIZ baseline: problem statement → generalized contradiction/principles → multiple solution report.
  5. Unstructured LLM baseline: same source and target context without archetype/mechanism scaffolding.
  6. Human technology-push baseline: expert teams identify and validate need–solution pairs under comparable time and evidence budgets.

Measurements that would make the distinction empirical

  • Hold out whole target domains and source archetypes, not merely documents, to test structural transfer rather than memorized case recombination.
  • Separate retrieval, mapping, problem discovery, proposal quality, and scrutiny effects through ablations.
  • Measure mechanism preservation independently of semantic similarity and domain distance.
  • Compare matrix coverage and yield against adaptive search; a fixed matrix may improve auditability but waste effort on implausible cells.
  • Run external prior-art search only after proposal generation and record whether it changes novelty or terminal disposition.
  • Blind critics to condition and source identity where possible; use critics not trained on the generator's positive examples.
  • Distinguish a generator's self-consistency, same-family relevance classifier, domain-expert review, adopter evidence, and actual empirical validation.
  • Preserve rejected candidates and repair trajectories so survivorship bias can be measured.
  • Report model versions, prompts, retrieved evidence, token/tool/time costs, and human labor.
  • Treat synthetic training use as a separate claim requiring held-out-domain or held-out-relation evaluation and contamination controls.

Limitations of this review

This is a scoping review conducted through general web and publisher search, not a database-complete systematic review. I did not have subscription-index coverage equivalent to Scopus, Web of Science, IEEE Xplore full text, ProQuest, Derwent Innovation, or commercial TRIZ/patent platforms. Paywalls limited inspection of some methods and appendices. Search was predominantly in English, while important TRIZ and technology-opportunity work exists in Russian, Chinese, Korean, and other languages. Terminology is fragmented across design cognition, engineering design, innovation management, patent analytics, biomimetics, HCI, and computational creativity. Commercial and unpublished systems may combine more stages than their public documentation reveals.

The comparison records only what the reviewed sources demonstrate or report; N and NR do not prove a capability is absent. Recent LLM systems and 2025–2026 publications may not yet have independent replications. “Complete proposal,” “independent criticism,” and “preserved trajectory” are operationalized from the review protocol and may not match authors' terminology. No historical-priority conclusion should be drawn from this memo.

Linked bibliography

Foundational analogy, cases, and representations

TRIZ, bio-inspired design, and problem evolution

Patent analogy and technology-opportunity discovery

Conceptual blending, computational creativity, and LLM systems