GM-TRIAD: Conditional Three-Channel Ground-Motion Generation with Native-Resolution Cascade Diffusion

Nonlinear time-history analysis requires motions with appropriate spectral and temporal characteristics, but recorded databases can be sparse under specified scenarios. GM-TRIAD generates three-channel ground motions through band-specific variational autoencoders and a native-resolution low–mid–high diffusion cascade, conditioned on scenario variables, response spectra and temporal-energy descriptors. With bounded record scaling, 1,188 of 1,245 design-spectrum combinations meet the 95–105% suite-mean spectral bounds. In a six-seed test, the uncorrected cascade meets the rule in 115 of 180 cases, against 54 for a joint single-resolution model. A controlled comparison with recorded-seed matching (REQPY) and random-phase synthesis uses equal suite sizes in two scenarios and 342 analyses of six steel and reinforced-concrete models. Median GM-TRIAD/REQPY ratios are 0.998 for peak interstory drift and 0.988 for peak floor acceleration. Mean floor cumulative absolute velocity is lower in all twelve scenario/model cells, with a median ratio of 0.847, alongside shorter input durations with a median ratio of 0.728. GM-TRIAD durations are 0.93 of the unmodified seeds; matching lengthens the seeds by factors of 1.18–1.83. In two further scenarios, analyzed descriptively, the cumulative-demand ratio is 0.685. Suite-median durations lie within one standard deviation of an empirical-model median in all fifteen suites, with the lowest ratios, 0.73‑0.86, at magnitude 7.5. In held-out tests, median generated/recorded duration ratios after correction are 1.15 under the scenario-and-spectrum conditions used for these suites and 0.98 with temporal-energy conditions. Within the NGA-West2 shallow-crustal domain, GM-TRIAD supplies additional candidates subject to spectral screening and scenario-based duration checks. Transfer to other tectonic settings remains unverified.