Answers are grounded in this paper and cite their sources.
The EASA AI Roadmap 2.0 establishes trustworthiness as the backbone of AI readiness in aviation, requiring that machine learning is not only accurate, but also deterministic, traceable to formally defined concepts, bounded by explicit safety constraints, and reconstructable post-hoc. Existing approaches treat machine learning as an unconstrained predictor over a latent feature space and apply constraints only as post-hoc filters. The central methodological problem remains open: a formal mapping from heterogeneous subsymbolic training evidence to admissible scenario adaptations that preserves these properties by construction. This paper formulates and realizes such a mapping, decomposable into formally defined substeps that satisfy four structural properties: determinism, semantic grounding, constraint compliance, and reconstructability. A realization for cyber-physical scenario-based flight training is presented. Its core is a deterministic conceptualized state space based on Deep-Linking Keywords that links theory-based assessment and simulator-based performance. A deterministic finite automaton governs instructional progression. Scenario adaptation is approximated by a constrained machine learning policy within an admissible action space enforced by monotonicity constraints, phase gating, and a projection operator that clips outputs to the admissible action space. The architecture is analytically verified by construction against the four properties and demonstrated empirically on N = 160 policy cases using k = 5 group-wise cross-validation. The constrained machine learning component approximates the deterministic baseline policy with a profile Mean Absolute Error (MAE) of 0.126 ± 0.024 and a $$+0.344$$ improvement in scenario-focus classification accuracy over the majority-class baseline. Statistical generalization beyond this proof-of-concept scope is identified as the next step.
No takes yet. Share an insight, caveat, or question.
Lilla et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: