Skip to Content
  • Proceeding Paper
  • Open Access

30 October 2026

6 Pages

Backward Causal Inference on Nuclear Knowledge Graphs: Domain-Specific Entity Resolution, Fault Path Discovery, and Blind Accident Prediction †

Independent Researcher, 02826 Görlitz, Germany
†
Presented at the 8th International Conference on Nuclear and Renewable Energy Resources (NURER 2026), Almaty, Kazakhstan, 10–12 September 2026.

Abstract

Probabilistic risk assessment (PRA) relies on manually constructed fault trees that cannot discover unknown root causes from observed failures. We present a causal knowledge graph approach for nuclear safety that addresses this limitation through backward transitive inference. The knowledge graph encodes 240 nuclear domain triplets spanning decay chains, reactor physics, material degradation, and safety systems with 11 domain-specific inference rules. A nuclear entity resolver replaces generic Jaro–Winkler string matching, achieving 97.8% accuracy on 1190 isotope notation pairs versus 19.9% for Jaro–Winkler, while reducing inference complexity from O(N2) to O(N1.3)—a 373× speedup with higher precision. Applied to the Davis–Besse reactor vessel head degradation, backward inference from 19 triplets enumerates 71 causal paths with 17 unique root causes. In a blind prediction test using only general pressurized water reactor (PWR) domain knowledge without accident-specific information, the system recovers four of five known Three Mile Island Unit 2 (TMI-2) root causes (80%). Validation against 1512 NUBASE2020 nuclides confirms 497/497 inferred decay chains with zero false positives. Convergence hub analysis identifies 33 common-cause failure candidates, and a hybrid confidence model—multiplicative for sequential chains, Möbius velocity addition for hub aggregation—is validated experimentally.

Article Metrics

Citations

Article Access Statistics

Multiple requests from the same IP address are counted as one view.