The methodology employs a topic-driven knowledge synthesis approach that combines semantic document embeddings, predefined topic taxonomies, and language model-powered text generation to extract structured best practices from scientific literature.

Core Approach

Unlike unsupervised clustering methods that may produce semantically incoherent groupings, this approach uses a predefined circular economy taxonomy to ensure interpretable and domain-relevant categorization. The methodology consists of four key innovations:

1. Predefined topic taxonomy: Documents are matched to 16 expert-defined circular economy topics using semantic keyword analysis, ensuring consistent categorization aligned with established CE frameworks
2. Anchor-context synthesis model: Within each topic, high-impact documents are selected as "anchors" and combined with semantically similar supporting documents to provide context for synthesis
3. Diversity-aware sampling: Anchor selection uses embedding-based diversity sampling to ensure coverage of different approaches within each topic
4. Multi-dimensional impact scoring: Each best practice is assessed across four dimensions (citation impact, innovation impact, commercial viability, replication potential)

Processing Pipeline

Figure 1: High-level overview of the best practice extraction pipeline

The platform leverages comprehensive scientific databases to filter the global knowledge base in natural sciences and engineering, identifying research that effectively applies AI methodologies to circular economy challenges. This multi-source approach ensures broad coverage of both academic publications and technological innovations.

Knowledge Base Filtering

The filtering process systematically narrows the global corpus to identify research at the AI-CE intersection:

1. AI methodology detection: Documents are scanned for AI/ML method keywords across six primary categories (machine learning, computer vision, NLP, optimization, prediction/forecasting, robotics)
2. CE topic matching: Semantic matching against 16 predefined circular economy topics ensures domain relevance
3. Quality filtering: Citation metrics, publication venue, and completeness criteria ensure high-quality source documents

Semantic Embedding

Documents are transformed into 768-dimensional semantic representations using SPECTER2, a transformer-based embedding model trained specifically on scientific documents. SPECTER2 captures semantic meaning beyond keyword matching, enabling:

• Computation of semantic similarity between documents
• Identification of supporting documents for context enrichment
• Diversity-aware sampling to ensure representative anchor selection

Embeddings are generated from the concatenation of document titles and abstracts, with results cached for computational efficiency across pipeline runs.

The platform uses a predefined taxonomy of 16 circular economy topics, each defined with associated keywords for semantic matching. This taxonomy-driven approach ensures consistent, interpretable categorization aligned with established circular economy frameworks and enables meaningful comparison across research areas.

Topic Matching Algorithm

Documents are assigned to topics using a weighted keyword matching algorithm that considers both direct keyword matches (higher weight) and semantic overlap with topic descriptions (lower weight). Each document is assigned to the topic with the highest matching score.

The synthesis process distills knowledge from topic-grouped documents into structured best practices using an anchor-context model with language model-powered text generation.

Synthesis Process

Synthesis Quality

Generated content is designed to follow academic writing conventions:

Generic terms such as "innovative," "cutting-edge," and "smart" are explicitly avoided to maintain scientific precision.

Each best practice is assessed across four complementary dimensions to provide a comprehensive view of its potential value. All scores are normalized to a 0-1 scale.

The Overall Impact Score is computed as the arithmetic mean of all four dimensions, providing a balanced assessment that does not unduly favor any single aspect.

AI methods are classified into six primary categories using a hybrid approach that combines keyword-based matching with language model classification for ambiguous cases.

Each best practice is assigned a primary AI method and may include secondary methods when multiple techniques are employed. The classification uses an expanded keyword taxonomy with weighted matching, falling back to language model classification when keyword matches are inconclusive.

The web-based platform provides multiple complementary views for exploring best practices, designed to support researchers, educators, and practitioners at participating HEIs.

Filtering Dimensions

The platform supports multi-dimensional filtering to enable focused exploration:

• AI Method: Machine Learning, Computer Vision, NLP, Optimization, Prediction/Forecasting, Robotics
• CE Topic: 16 predefined circular economy domains
• Document Type: Paper-based, Patent-based, or Hybrid practices
• Technology Readiness Level: Estimated TRL (1-9)
• Publication Year: Temporal filtering
• Impact Score: Minimum threshold filtering
• Geographic Location: Country and jurisdiction-based filtering

Current Limitations

• Topic coverage: The predefined taxonomy may not capture emerging or niche CE topics not represented in the current 16 categories
• Language bias: The corpus is primarily English-language, potentially missing relevant research published in other languages
• TRL estimation: Technology Readiness Levels are estimated from text analysis rather than direct assessment, introducing uncertainty
• Citation lag: Recently published documents may have artificially low citation impact scores due to citation accumulation delay
• Synthesis quality: Language model-generated content may occasionally introduce factual errors or overstate findings

Future Directions

• Dynamic taxonomy: Automatic discovery of emerging topics from document clusters
• Cross-lingual analysis: Extension to non-English literature
• Validation framework: Expert review and quality scoring of synthesized practices
• Real-time updates: Continuous ingestion of new publications
• Citation network analysis: Identifying influential research trajectories
• SciSciNet integration: Additional data source for enhanced bibliometric analysis