Cross-domain transfer of seismo-volcanic classifiers across geologically distinct volcanic systems remains a critical challenge: models trained on one volcano degrade when deployed at a new site without target-domain labels. This work proposes the Phase-Modulated Hypergraph Fusion (PMHF) model, a trimodal deep architecture that jointly exploits CLIP-encoded spectrograms, waveform images, and physics-informed seismic features to discriminate Long-Period (LP) from Volcano-Tectonic (VT) events in an adaptation domain setting. PMHF integrates three novel modules: a Spectral Calibrated Physics Gate (SCPG) that recalibrates tabular features via class-conditional resonance matrices; Dynamic Hypergraph Convolutions (DHGC) that propagate visual features through probability-conditioned hyperedges; and a Trimodal Episodic Cross-Attention (TECA) mechanism that fuses all pairwise modal interactions through persistent memory. Three leave-one-volcano-out scenarios evaluated across Galeras, Cumbal, and Chiles-Cerro Negro (bootstrap N=1000) show that PMHF achieves mean balanced accuracy 0.765, peaking at 0.868±0.009 (MCC =0.743±0.017) for Galeras+Cumbal→Chiles-Cerro Negro, where bimodal baselines plateau below 0.659 and tabular models exhibit systematic class inversion (MCC =−0.405). Domain adaptation diagnostics confirm that trimodal fusion achieves the tightest and most stable source-target feature alignment, enabling robust volcanic monitoring without retraining.

Reliable discrimination of Volcano-Tectonic (VT) and Long-Period (LP) seismic signals is central to volcanic surveillance, yet operational deployment is hindered by pronounced inter-volcano domain shift, scarce annotated catalogues, and heterogeneous acquisition instrumentation. We present a self-supervised trimodal deep learning framework that jointly encodes spectrograms, raw one-dimensional waveforms, and physical tabular features, and we evaluate its capacity for cross-domain generaliza-tion across three leave-one-volcano-out transfer settings among Galeras, Cumbal, and Chiles-Cerro Negro (Colombia). Beyond classification accuracy, we systematically characterise the geometry of the learned latent space using class- and domain-separability landscapes projected into two dimensions, quantifying silhouette coefficients, Davies–Bouldin indices, separation indices, and decision-boundary alignment for every ablation variant. The trimodal framework (SSL Trimodal) attains the highest bootstrap balanced accuracy on the Galeras+Cumbal→Chiles-Cerro Negro transfer (0.9080±0.0072) and the most favourable class-separation index (Sep-Idx_2D = 2.06) among all eight evaluated variants in that scenario. It further achieves the best class-separation index (1.72) even on the most divergent transfer setting, Galeras+Chiles-Cerro Negro→Cumbal, despite not attaining the highest accuracy there. This dissociation between point accuracy and embedding quality is, we argue, the more in-formative signal for assessing cross-domain robustness than accuracy alone. We report this result alongside its limitations transparently: the CORAL covariance-alignment baseline outperforms the trimodal framework on two transfer settings with comparable mean accuracies ( 0.81), underscoring that discriminative contrastive pretraining and second-order domain alignment address complemen-tary, rather than redundant, failure modes. To our knowledge, this is the first trimodal self-supervised framework reported for cross-domain generalization in seismo-volcanic signal classification.