The safe operation of nuclear power plants requires diagnostic systems capable of identifying not only individual accident types but also concurrent multi-accident compositions—scenarios in which two or more abnormal events occur simultaneously. Historical incidents such as Three Mile Island and Fukushima Daiichi have demonstrated that cascading and compound failures can produce plant responses that differ qualitatively from any individual event. However, developing machine learning-based diagnostic systems for multi-accident scenarios faces a fundamental combinatorial bottleneck: for N individual accident types, the number of pairwise compositions grows as N(N−1)/2, and each composition conventionally requires its own set of high-fidelity coupled thermal-hydraulic simulations and dedicated training data. This data generation burden is prohibitively expensive and, for large accident taxonomies, practically infeasible.
This paper presents the Concept-Vocabulary Compositionality (CVC) framework, which achieves zero-shot diagnosis of multi-accident compositions using only single-accident training data. The key insight is that vector quantization transforms continuous sensor representations into discrete concept assignments, producing concept histograms that summarize each simulation run as a distribution over a fixed vocabulary. Because the concepts are discrete and countable, the concept histogram of a multi-accident run can be naturally understood as the superposition of the concept histograms of its constituent single-accident types.
The CVC framework decomposes the reactor's 113 operator instrument channels into six physical subsystem groups based on the functional organization of pressurized water reactor instrumentation and assigns each subsystem its own 32-entry vector quantization codebook. This produces a vocabulary of 192 discrete concepts that capture subsystem-specific transient patterns. For composition diagnosis, we introduce pairwise differential scoring: a zero-shot method that identifies which pair of concurrent accidents best explains an observed concept histogram by comparing against the additive superposition of single-accident prototypes. No multi-accident data is used at any stage of training or prototype construction.
We evaluate the framework on the Compact Nuclear Simulator (CNS) for a Westinghouse-type 600 MWe PWR. The model trains on 400 single-accident runs covering three accident types—Loss of Coolant Accident (LOCA), Steam Generator Tube Rupture (SGTR), and Main Steam Line Break (MSLB)—deliberately selected as a stress test because all involve primary–secondary pressure boundary breaches with heavily overlapping subsystem signatures. Testing on 500 genuinely coupled multi-accident simulation runs across three pairwise compositions, the best seed-ensembled scoring method achieves 97.0% zero-shot composition diagnosis accuracy (Clopper–Pearson 95% CI: [96.0%, 97.8%]). A systematic ablation establishes that the discrete concept vocabulary provides statistically significant advantages over continuous features (Wilcoxon signed-rank p = 0.008, Cohen's d = 0.99), with the ensemble gap reaching 17.9 percentage points due to codebook convergence enabling constructive score averaging.
Beyond accuracy, concept fingerprint analysis reveals that each accident type develops physically interpretable concept signatures in distinct reactor subsystems, and these fingerprints persist in multi-accident compositions through selective override. This compositional transparency directly addresses the interpretability requirements for regulatory acceptance of ML-based diagnostic systems in nuclear safety applications. The framework eliminates the combinatorial simulation data burden and enables diagnostic systems to handle unforeseen concurrent failures from the outset—providing inherent generalization capability that supports probabilistic safety assessment and risk-informed decision-making.