Seismic events classification using language syntax and biomimesis

The planet’s surface is a restless orator: its fractures, failures and flows in response to climatic, tectonic and anthropogenic forcing inscribe narratives in seismic waveforms. Our limited success in decoding these narratives by classification and attribution of complex, incognito signals, threatens to leave Environmental Seismology data-rich but epistemologically impoverished. We propose that geomorphic signals can be formalized as language and tracked as evolving lineages. Hence, we develop a self-organizing, classification method with comprehension, resulting in an explainable, evolving phylogenetic tree. Unlike supervised classification methods that require exhaustive labels or clustering algorithms that conflate statistical similarity with physical kinship, our classification tool learns without labels, generalizes without forgetting physics, and explains without obfuscation.

We present a “glass-box” classifier, unsupervised in perception but supervised in its definition, that treats seismic data not as flat feature vectors but as structured, generative text, translating ground motion into a lexicon of geomorphic processes. Our system discovers its own alphabet, syntax, and semantics: autonomously constructing a taxonomic tree for seismic events while remaining interpretable. To do so, we break down the continuous seismic signal into discrete "phonemes." Multi-scale temporal descriptors, impulsive micro-textures (1.25 s), meso-scale envelope dynamics (5 s) and slow background trends (20 s), compose a multi-metric feature suite. These windows are fused into a tensor encoding nonlinear force interactions, then discretized through RVQ into context-aware symbols. Thus, we replace the geometric rigidity of static clustering with a dynamic evolutionary state space where signal classes behave as adapted species (rockfalls, landslides, debris flows, mine collapses, volcanic tremors, GLOFs, tectonic and glacial earthquakes, nuclear explosions, anthropogenic noise), governed by the Free Energy Principle. Similarity of process signals is tripartite: syntactic (grammar divergence), information-theoretic (surprisal), and algorithmic (NCD). This distinguishes events that look alike but have differing mechanisms: a debris flow and lahar may share "rumble" words, yet obey distinct physical grammars.

Modeled on Darwinian phylodynamics, the spectral species defined by their grammatical structure (causality), "metabolize" incoming data by minimizing thermodynamic surprise. The populations undergo sympatric speciation, hybridization, commensalism, and extinction, disentangling the "phylogenetic distance" (similarity) between superficially similar signals, and ultimately resulting in optimized classification. The algorithmic biomimicry of our approach outperforms static taxonomies that fail in non-stationary Earth systems without retraining.

Applied to a global, geologically heterogeneous inventory of >6000 curated records, preliminary results show phonemes reliability reaches 94–98% across stations and a >50% drop in articulation-structure complexity from noise to geomorphic events. The inferred phylogeny is physically meaningful, decoupling categories in distinct topological manifolds, allowing the classifier to reject false positives without supervision. Uncertainty is metabolized: high-aleatoric/low-epistemic signals (inherent noise) separate from low-aleatoric/high-epistemic anomalies (black swans: candidate new species). Free Energy scores, combine complexity and inaccuracy, outperform baselines in robustness to gaps, clipping, and dropout by over 20%.

The model is self-interpreting; rather than an opaque class label, it outputs a semantic sentence, exposing the decision path to a physically meaningful event classification, whilst also encapsulating information unique to specific events. Legible, self-improving classification turns detection into naming, and identity into process understanding.