Machine Analysis of Global Rock Records Reveals the Permian Extinction Was a Multi-Driven Collapse

Roughly 252 million years ago, at the boundary between the Permian and Triassic periods, life on Earth came closer to total annihilation than at any other moment in its history. An estimated 80 to 90 percent of marine species vanished, alongside a substantial share of terrestrial vertebrates and plants, and ecosystems needed millions of years to recover. Scientists have long debated which environmental killer—or combination of killers—drove this catastrophe. A new systematic review published in Frontiers in Earth Science takes an unusually quantitative approach to that question, compiling paleoenvironmental data from dozens of studies worldwide and crunching them with multivariate statistics. Its conclusion: the end-Permian crisis was not a single-cause event but a multidimensional environmental system in which ocean chemistry, climate, volcanism, and local geology each operated largely on their own terms.

The research team, led by Atif Salim and Hanlie Hong of the China University of Geosciences in Wuhan, together with Thomas J. Algeo of the University of Cincinnati, systematically gathered and standardized multi-proxy paleoenvironmental data from 40 peer-reviewed studies covering 38 Permian-Triassic Boundary (PTB) sections around the world. Chinese sections, which include the world-famous Meishan, Shangsi, and Xiakou profiles, accounted for 57.5 percent of the compiled studies, reflecting the exceptional stratigraphic resolution of the South Chinese Tethyan record. The remaining 42.5 percent came from international localities spanning the Tethyan, peri-Gondwanan, and Panthalassic paleogeographic realms.

From these studies the team built a combined dataset spanning at least eight major proxy categories: stable isotopes of carbon, oxygen, and sulfur; redox-sensitive trace elements such as molybdenum, uranium, and vanadium; rare earth elements; sedimentological and lithofacies observations; paleoclimatic indicators including the chemical index of alteration and conodont oxygen isotopes; ocean chemistry markers of anoxia, euxinia, and acidification; paleobiological indicators such as microbialites, conodonts, and foraminifera; and volcanogenic signals like mercury anomalies and ash layers. Each observation was linked to its published source, section name, stratigraphic position, and sampling resolution, so that every statistical result could be traced back to the underlying rock record.

The analytical engine of the study combined two computational techniques. First, the team applied Principal Component Analysis, or PCA, an unsupervised dimensionality-reduction method that transforms many correlated variables into a smaller set of orthogonal components. All numerical variables were z-score normalized beforehand, giving every proxy a mean of zero and a standard deviation of one so that no single measurement type dominated the results simply because of its units. Components were retained using the Kaiser criterion, which keeps only those with eigenvalues greater than one, supplemented by inspection of a scree plot. Second, the researchers constructed a correlation-based network using the NetworkX package in Python, treating each proxy category as a node and statistically significant Pearson correlations as edges, then measuring connectivity and edge density to quantify how tightly the different environmental processes were coupled.

The PCA delivered a striking headline result: four principal components together explained 71.2 percent of the total variance in the dataset. The first component, accounting for 24.3 percent of the variance, was defined by strong positive loadings of ocean chemistry (0.418), trace elements (0.391), and stable isotopes (0.342). In plain terms, this axis captures the geochemical-redox gradient of the crisis—the widespread development of anoxic and euxinic seafloor conditions, recorded by enrichments of redox-sensitive metals and pyrite framboid distributions, coinciding with the large negative carbon isotope excursions that signal massive perturbation of the global carbon cycle. The co-loading of redox proxies with carbon isotopes suggests that carbon cycle dynamics and ocean chemistry were linked during the extinction interval, consistent with earlier work on South Chinese sections.

The second component, explaining 19.2 percent of the variance, loaded most heavily on rare earth elements (0.412), paleoclimatic indicators (0.361), and paleobiological indicators (0.349). This axis represents surface-system processes: continental weathering intensity, climate variability, and the biological responses tied to them. Rare earth elements track terrestrial input and weathering, while proxies such as conodont oxygen isotopes record temperature. The third component, at 15.8 percent of variance, paired paleoclimatic indicators (0.512) with volcanogenic indicators (0.412) and stable isotopes (0.428)—a clear volcanism-climate coupling signal pointing to the Siberian Traps, the vast volcanic province whose eruptions of carbon dioxide, methane, and sulfur gases are widely regarded as a principal trigger of rapid warming and ocean acidification. The fourth component, contributing 11.9 percent, was dominated by sedimentological data (0.489), indicating that local depositional factors—lithofacies, sedimentary structures, and basin-specific processes—varied partly independently of the global geochemical and climatic signals.

Perhaps the most provocative finding came from the network analysis. When the team filtered the correlation network at the conventional strong-correlation threshold of |r| greater than 0.5, no edges survived. None of the proxy relationships proved strong enough to qualify. The authors are careful to note that this does not prove the PTB’s environmental processes were truly independent—Pearson correlation cannot capture nonlinear relationships or temporal lags between processes—but the weak connectivity does suggest that ocean redox change, carbon cycling, weathering, volcanism, and biological response operated as loosely coupled, partly parallel strands rather than as one tightly synchronized cascade. Each proxy, in other words, captures a different facet of the Earth system, and only by combining them can a fuller picture emerge.

The study also exposed uncomfortable biases in the existing literature. Sedimentological and lithofacies data appeared in 95 percent of the compiled studies and ocean chemistry indicators in 77.5 percent, but volcanogenic indicators showed up in only 40 percent, and rare earth elements in a mere 15 percent. That underrepresentation matters enormously, because volcanogenic proxies are the very tools needed to test the Siberian Traps hypothesis and quantify volcanic forcing, while rare earth elements constrain weathering dynamics on land. Geographically, the analysis revealed clear clustering of Tethyan, especially Chinese, sections, with weak dispersion of Gondwanan and Panthalassic records—meaning global interpretations of the extinction have been disproportionately shaped by one paleogeographic realm.

The authors argue that these results reframe the PTB not as a single-cause extinction but as a multi-causal, cascading event with feedback loops, challenging older reductionist models that emphasized one trigger. They also caution that the observed weak coupling may partly reflect methodological inconsistency across studies—differences in analytical techniques, sampling resolution, and proxy selection—rather than purely environmental complexity. The PCA and network patterns, they stress, should be read as descriptions of the available standardized dataset, not as fully representative estimates of all global PTB environments.

Looking forward, the team lays out a clear research agenda: expand sampling into underrepresented peri-Gondwanan, Panthalassic, and high-latitude regions; prioritize volcanogenic proxies and rare earth element measurements; and adopt standardized data collection and reporting protocols so that future datasets can be integrated seamlessly. Where sufficiently large and well-resolved datasets become available, more advanced statistical and computational methods, including predictive machine learning, could complement the multivariate framework demonstrated here. For now, the message is clear: the greatest extinction in Earth’s history cannot be understood through any single environmental signal, and the rocks that recorded it still hold far more stories than any one proxy can tell.

Subject of Research: Multi-proxy statistical analysis of paleoenvironmental change across Permian-Triassic Boundary sections

Article Title: Integrating multi-proxy records using multivariate statistical and computational approaches: a systematic review of paleoenvironmental variability across Chinese and global PTB sections

Article References: Integrating multi-proxy records using multivariate statistical and computational approaches: a systematic review of paleoenvironmental variability across Chinese and global PTB sections. (n.d.). https://doi.org/10.3389/feart.2026.1908791

Image Credits: AI Generated

DOI: 10.3389/feart.2026.1908791

Keywords: Permian-Triassic Boundary, mass extinction, paleoenvironment, principal component analysis, network analysis, Siberian Traps, ocean anoxia, geochemical proxies, South China, paleoclimate, systematic review, volcanism