Bridging research and policy in China's energy sector: A semantic and reinforcement learning framework

Aligning academic research with policymaking is vital for addressing China's energy challenges. This study introduces an AI-driven framework combining BERTopic, semantic similarity analysis, and deep reinforcement learning (DRL) to evaluate alignment between 106,661 English-language academic papers and 618 national-level policy documents. Topic modeling reveals strong convergence in themes such as “Coal mining and geological formations,” which account for 15.73 % of academic publications, while “Safety regulations and worker protection” dominate policy texts at 13.35 %. In contrast, emerging topics like “Digital economy and carbon transformation” remain underrepresented, with a popularity score of only 0.13. Semantic similarity analysis across 22 policy and 27 academic topics yields an average cosine similarity of 0.23, with only 12.5 % of topic pairs exceeding 0.4, underscoring thematic misalignment. Structurally, policy networks are 15.9 times denser and exhibit 30 × higher clustering coefficients than scientific networks, indicating more centralized but less diversified discourse. DRL-based prioritization identifies “Power systems and renewable integration” as the top-performing theme (Q-value = 1.6225), highlighting opportunities for targeted energy transition policies. These quantitative results offer empirical evidence to guide theme-based policy adaptation and foster actionable science-policy integration. • Integrates AI methods to assess research-policy alignment in energy governance. • Identifies thematic gaps in China's energy policy using NLP and deep learning. • Uses deep reinforcement learning for dynamic policy prioritization. • Demonstrates framework scalability across environmental and policy domains. • Provides actionable insights for evidence-based energy policy decisions.