Background

Science policies increasingly emphasize open access (OA) as a mechanism for enhancing the circulation and societal impact of scientific knowledge. Lowering barriers to access, OA is expected to facilitate the reuse of research beyond academia, including in technological innovation. One domain where this relationship can be observed is the interaction between scientific publications and patents. Patents frequently cite scientific literature through non-patent references (NPRs), which have long been used as indicators of sciencée–technology linkages.

Despite the expansion of open access, relatively little is known about how different OA publishing models shape the ways in which scientific knowledge is mobilized in patents. Existing studies typically focus on whether open access publications are more likely to be cited in patents, often treating OA as a binary characteristic (Jahn et al., 2022; Ottonicar et al., 2020). However, open access encompasses a wide range of publishing arrangements, such as hybrid, bronze, green, gold, and diamond OA, that differ substantially in governance structures, business models, and integration into scholarly communication infrastructures.

Moreover, the presence of a citation in a patent does not necessarily indicate substantive use of scientific knowledge. Patent citations may reflect legal disclosure requirements, examiner-added prior art, or routine search practices rather than direct engagement with scientific research (Bryan et al., 2020; Marx & Fuegi, 2022). As a result, citation frequencies alone provide only a partial understanding of the role of open access in technological innovation.

Research questions

This paper proposes to distinguish between two processes through which scientific knowledge becomes incorporated into technological invention: selection and translation. In the selection stage, scientific publications must first be “discovered” and retrieved by inventors or patent examiners within the broader landscape of scientific information. In the translation stage, the selected knowledge is integrated into the technological content of the patent, implying a degree of conceptual alignment between scientific research and the invention.

Building on this perspective, the study addresses two main questions:

• How does the distribution of open access models among scientific publications cited in patents compare to the distribution observed in the scientific literature as a whole?
• Conditional on being cited in patents, do publications disseminated through different OA models differ in their cognitive proximity to patented technologies?

Data and methods

The empirical analysis relies on large-scale data linking patents to scientific publications through non-patent references (NPRs). We use the global dataset developed by Marx and Fuegi, which contains more than 44 million patent–publication references extracted from patent documents worldwide (Marx, 2023). Each reference links a patent to a scientific publication cited as prior art and provides information about the location of the citation within the patent document (front section, body of the patent text, or both).

These NPRs are matched with publication metadata from OpenAlex, which provides information on publication year, disciplinary classification, citation impact, and open access status (closed, hybrid, bronze, green, gold, and diamond OA). Patent abstracts are retrieved from the PatentsView database (https://patentsview.org/home), allowing textual comparison between patents and the scientific publications they cite.

Because the full NPR dataset contains tens of millions of observations, we construct analytical samples that allow balanced comparison across citation positions. Following robustness checks, we draw random samples of 50,000 publication–patent pairs for each citation position (front citations, body citations, and citations appearing in both locations), resulting in three comparable datasets.
To evaluate the cognitive relationship between patents and scientific publications, we compute semantic similarity between patent abstracts and the abstracts of cited publications. Text embeddings are generated using the SPECTER-2 language model (https://github.com/allenai/SPECTER2), a transformer-based model trained on citation networks that maps scientific documents into a shared semantic space. For each patent–publication pair, cosine similarity between embedding vectors is used as an indicator of conceptual proximity between scientific knowledge and technological invention.

This approach makes it possible to distinguish between the selection of publications into patent citations and the degree of cognitive alignment between cited science and patented technologies, while accounting for differences across open access models and citation positions within patents.

Results

The results reveal a clear distinction between visibility-driven selection and substantive knowledge integration.
First, the distribution of open access models among patent-cited publications differs markedly from the distribution observed in the scientific literature. Hybrid and bronze OA publications are significantly overrepresented among patent citations across disciplinary domains. These models are often associated with prestigious journals hosted on large commercial publishing platforms and benefit from strong integration into indexing services and search tools used in patent offices.

Second, the semantic analysis shows that this dominance in citation frequency does not correspond to stronger cognitive alignment with patented technologies. When semantic similarity between patents and cited publications is examined, publications disseminated through fully open access models, particularly gold OA journals and, to a lesser extent, diamond OA journals, exhibit equal or higher levels of semantic proximity to patented technologies.

These differences are especially pronounced for citations appearing in the body of patents, which are more likely to reflect substantive engagement with scientific knowledge during the inventive process. In contrast, differences across publishing models are weaker for front-section citations, which are often associated with formal disclosure practices.

Discussion and implications

To sum up, these findings suggest that patent citation patterns reflect not only technological relevance but also the institutional organization of scholarly communication and information infrastructures. Publishing models that are strongly embedded in dominant discovery systems benefit from greater visibility and are therefore more likely to be retrieved and cited during patent preparation and examination.

At the same time, fully open access journals may disseminate research that is equally or more relevant to technological problem-solving, yet remain less visible within these infrastructures. This result highlights the importance of discoverability as a key mechanism linking scholarly communication systems to technological innovation, and suggests that patent citation patterns may, to some extent, reproduce the prestige-based hierarchies observed in scholarly publishing.

More broadly, the study suggests that the impact of open access on innovation cannot be understood solely through access to publications. The way in which publishing models are embedded in scholarly information systems, through metadata standards, indexing practices, and discovery platforms, plays a central role in shaping how scientific knowledge is mobilized in technological development.
Improving the discoverability and integration of fully open publications within innovation-relevant information systems may therefore be as important as expanding open access mandates themselves.