A bibliometric mapping of research trends in artificial intelligence-enabled virtual reality STEM education

Synthesis: Ajimotokan and colleagues (2026) map the research literature on AI-enabled virtual reality in STEM education, drawing 241 publications from Dimensions.ai for 2005–2025 and screening them through PRISMA 2020. The unit of analysis is publications, not learners: no outcomes were collected or pooled, so nothing here tests whether AI-enabled VR improves learning. The descriptive pattern is stark. 204 publications (84.6%) appeared between 2023 and 2025, including 112 in 2025 alone, and the growth inflection falls in 2023, the first year annual output (29) exceeded the whole 2005–2019 cumulative total (14). Output was unchanged at two publications in both 2019 and 2020, which the authors read as evidence that COVID-19 did not drive the field; the timing instead coincides with Generative AI. Field-normalized citation performance is strong, with 92 of 127 publications (72.4%) at or above the world average, yet authorship and venues are extremely dispersed. The distinction worth keeping: a fast-growing, well-cited, unconsolidated field.

Key Findings

1. The Dimensions.ai search returned 1,391 records; period, document-type and classification filters left 337, and eligibility screening removed 96, producing 241 publications from 2005 to 2025.
2. 204 of the 241 publications (84.6%) appeared in 2023–2025, including 112 in 2025 alone, while only 14 (5.8%) appeared between 2005 and 2019 and six years had none.
3. The growth inflection is 2023, when 29 publications exceeded the 2005–2019 cumulative total of 14; 2022 also recorded 14 and matched rather than exceeded that threshold.
4. Field-normalized citation performance is strong: median Field Citation Ratio 2.50, with 92 of 127 publications (72.4%) at or above the world average and a maximum of 64.03.
5. The corpus came from 789 authors, 774 of them (98.1%) with a single publication, and 209 venues, 185 (88.5%) with a single publication; the Gini coefficients were 0.021 and 0.117.
6. Term co-occurrence reduced 7,166 extracted terms to 101 in four clusters; the top terms by total link strength were engagement (449), educator (385), science (376), issue (365) and STEM education (362).
7. SDG 4 was assigned to 203 publications (84.2%) but SDG 9 to only 19 (7.9%), a gap the authors flag given the field's engineering and Industry 4.0 framing.

A recent corpus inside a twenty-year window

The corpus is narrower than its date range suggests. Only 14 publications predate 2020, and 225 of 241 (93.4%) appeared from 2021 onward. Conference proceedings supplied 92 publications (38.2%), and 174 (72.2%) addressed engineering, technical or vocational education explicitly. The inflection rule is stated plainly: the first year whose annual output exceeded the cumulative 2005–2019 total. That rule gives 2023 the inflection by a single publication over 2022, and the authors concede the threshold is sensitive to a small baseline. Their scope check on the 174 engineering-only papers moves the inflection to 2022, since that subset's baseline of 10 was exceeded by its 12 papers that year; they treat this as threshold sensitivity, because the largest acceleration still arrives after 2022 in both datasets. The pandemic claim rests on the same counts: two publications in 2019, two in 2020, and a merged COVID/pandemic term averaging 2023.6, which suggests retrospective discussion rather than a driver.

Strong normalized impact, dispersed structure

Raw citation counts mislead here, and the paper says so before presenting them. The 241 publications have 1,855 citations, a mean of 7.70 and a median of one, with 95 (39.4%) uncited and an h-index of 22. The 2025 cohort holds 112 publications (46.5% of the corpus) but only 256 citations (13.8%), a citation-window effect rather than a quality signal. Field Citation Ratio values exist for 127 publications (52.7%), because the metric is computed only for papers at least two years old, so the 2025 cohort drops out. Within that subset the median FCR is 2.50 and 21 publications (16.5%) exceed ten times expected impact. Normalization reorders the field: only five publications appear in both the raw-citation and FCR top ten, and the most-cited paper by raw count (129 citations) ranks fifth by FCR. Consolidation is absent. No author published more than three papers and no venue more than four, well beyond the dispersion Lotka's Law predicts, and multiple correspondence analysis returns one dominant cluster with two dimensions explaining 34.28% of variance.

What the map is organized around

The network's most connected terms are pedagogical rather than technical: engagement, educator, science, issue and STEM education. Explicit technology terms rank lower, which the authors read as a shift from demonstrating tools toward adoption questions. Four clusters cover adoption and accessibility, immersive technologies and performance, institutional and pedagogical transformation, and Industry 4.0 workforce competencies. Mixed reality is the most central term by distance from the map origin, while industrial training, high cost, mathematics, graduate and policymaker sit at the periphery. The recency signal is the sharpest finding: accessibility (23 occurrences, total link strength 248) averages 2024.7 and is among the ten most central terms, and ethical concern, policymaker, scalability, equity and data privacy carry the latest average years. Generative AI is present but thin, with 8 occurrences and a total link strength of 61, the lowest quintile, and an average year of 2024.8. Blockchain is absent after relevance filtering, and neuroadaptive systems do not appear at all. None of this measures effectiveness, only publication patterns.

What this means for practice

• Instructors. Read the primary studies before adopting a tool. The map counts papers and citations, not learning gains, and cannot establish whether AI-enabled VR works.
• Researchers. Build sustained programs and shared instruments rather than one-off studies, since 774 of 789 authors (98.1%) published once and no differentiated schools appear.
• Administrators. Fund evaluation alongside deployment. Cost, accessibility and privacy are recent, weakly connected terms, and the corpus offers no outcome evidence for procurement.
• Research support staff. Report affiliation coverage with any country ranking: author-country metadata exists for 159 of 241 publications (66.0%), corresponding-author country for 38 (15.8%).

Limitations

• Dimensions.ai is the sole source, so the distributions describe the Dimensions-indexed literature rather than the global literature, as the authors state.
• The export carried no cited-reference data, which precludes co-citation and bibliographic coupling analysis of intellectual structure.
• Field-normalized citation analysis covers 127 of 241 publications (52.7%) and excludes the entire 2025 cohort of 112 papers.
• Author-country metadata is missing for 82 publications, and those records are newer, more often open access and less cited, so country rankings would underrepresent recent work.
• Terms were mined from titles and abstracts because the export lacked author or indexer keywords, and VOSviewer map coordinates depend on normalization, layout parameters and random initialization; the analysis used Bibliometrix in R version 5.4.1.

Citation

Ajimotokan, H. A., Ajibade, Q. A., Saidu, N. S., Kareem, A. B., Otuoze, A. O., Sadiku, I. B., & Mahmud, H. O. (2026). A bibliometric mapping of research trends in artificial intelligence-enabled virtual reality STEM education. Discover Education, 5(1), Article 1305.