Discontinuous Cyclostratigraphic Age Models Recovering Using a Physics-Informed Reviewer Network Addressing Hidden Hiatuses in Astrochronology

Traditional cyclostratigraphic orbital tuning for geochronology is often formulated as a constructive problem: given paleoclimate proxy data in the depth domain, one seeks to generate a smooth and continuous age model that best matches a theoretical orbital forcing target. However, this constructive paradigm faces a fundamental limitation: it intrinsically favors visually continuous solutions. When hidden hiatuses—gaps in deposition without clear lithological or sedimentological evidence—are present in a sedimentary succession, the algorithm tends to silently stitch across missing time intervals by stretching or compressing the data excessively. While the resulting continuous age models may appear morphologically plausible, they can substantially violate global orbital phase consistency, leading to systematic biases in the tuned chronologies. To overcome this bottleneck, we propose a novel methodological framework for astrochronological tuning. Unlike traditional algorithms that directly construct a continuous age model, we first train a physics-informed reviewer network to assess the local self-consistency of candidate age models in the orbital-state space. The reviewer network is then used as a failure detector to identify the onset of age-model breakdown, which marks a candidate location of a hidden hiatus. Once a breakdown point is detected, the system abandons the continuity assumption and instead performs independent orbital tuning on the sedimentary segments above and below the candidate hiatus. Finally, the hiatus duration Δt is estimated from the boundary-age difference between the two independently tuned segments, yielding a segmented age model that reflects the true stratigraphic record rather than a falsely continuous one. Validation experiments on synthetic data show that the physics-informed reviewer network can accurately localize the onset of hidden discontinuities, whereas direct inversion of the hiatus duration Δt from reviewer scores or local continuity objectives yields unstable results. In contrast, the proposed segment-wise recovery, combined with boundary-age differencing, provides a more interpretable estimate of missing time and reconstructs the correct segmented temporal structure more reliably. This framework moves cyclostratigraphy beyond the traditional role of validating continuous age models and enables automated reconstruction of discontinuous age models with explicit consideration of the incompleteness of geological records.