Basalts are the most abundant rocks on the ocean floor, and their chemical fingerprints carry the memory of the mantle sources and geological processes that produced them. Mid-ocean ridge basalts, known as MORB, rise from the depleted upper mantle at spreading centers. Ocean island basalts, or OIB, tap enriched mantle plumes beneath volcanic islands such as Hawaii. Island arc basalts, abbreviated IAB, form above subduction zones, where descending oceanic slabs release fluids that modify the mantle wedge and introduce components from the crust. Telling these three families apart is fundamental to reconstructing plate tectonic histories, identifying ancient oceanic crust preserved in mountain belts, and understanding how the mantle has evolved over billions of years.

The classical approach uses discriminant diagrams, which plot ratios of two or three trace or major elements against one another and delineate fields corresponding to different tectonic settings. Geologists plot an unknown sample on the diagram and read off its likely origin from the field in which it falls. While elegant and easy to use, these diagrams have well-documented shortcomings. Their boundaries are drawn by eye around limited sample sets, the fields often overlap indistinctly, and a substantial fraction of real samples fall into ambiguous zones or outside the diagram’s coverage altogether. Previous large-scale audits of discrimination diagrams have found that misclassification rates can be uncomfortably high, and the new study quantifies that gap directly.

To build a more rigorous benchmark, the team turned to two of the largest open-access geochemical repositories in the Earth sciences: the GEOROC database and PetDB, maintained by EarthChem. From these archives they compiled 950 samples of MORB, OIB, and IAB, each characterized by a suite of major and trace element concentrations. This curated dataset formed the raw material for both a statistical exploration of the data and the training of machine learning classifiers.

The first analytical step was principal component analysis, a technique that compresses many correlated chemical variables into a small number of composite axes that capture the greatest variance in the data. The analysis revealed that the first principal component, PC1, is dominated by so-called enriched elements, most notably niobium, a trace element that behaves incompatibly during mantle melting and accumulates in melts derived from enriched mantle sources. The second principal component, PC2, corresponds mainly to silica dioxide and aluminum oxide, the major oxides that track the degree of melting and the mineralogical character of the source. Together, PC1 and PC2 explain more than 60 percent of the total chemical variance in the dataset, and, crucially, plots of samples in this two-dimensional space effectively separate the three basalt types. In other words, the essential information needed to distinguish MORB from OIB and IAB is genuinely present in the chemistry, and it can be extracted without hand-drawn boundaries.

With the feature structure established, the researchers trained three main classification models drawn from the standard machine learning toolbox: support vector machines, random forests, and k-nearest neighbors. They also included XGBoost, a gradient-boosting algorithm, as a supplementary benchmark. Each method approaches the problem differently. Support vector machines construct decision boundaries in a high-dimensional feature space, maximizing the margin between classes. K-nearest neighbors classifies a sample by asking which category dominates among its most chemically similar neighbors in the training set. Random forests, first formalized by Leo Breiman in 2001, build hundreds of decision trees, each trained on a random subset of the data and features, and then pool their votes. XGBoost instead builds trees sequentially, with each new tree trained to correct the errors of its predecessors.

Model evaluation followed best practices in machine learning. The team used five-fold cross-validation, in which the data are repeatedly split so that models are tested on samples they never saw during training, and they reported performance using confusion matrices and receiver operating characteristic curves, tools that capture not just overall accuracy but also how well each class is distinguished and how the trade-off between sensitivity and specificity behaves. Final performance figures were calculated on an independent test set, providing an unbiased estimate of real-world accuracy.

The results were striking. Random forests came out on top with an overall accuracy of approximately 0.97 on the independent test set. Support vector machines followed at about 0.95, the XGBoost benchmark reached roughly 0.86, and k-nearest neighbors trailed at about 0.83. For comparison, the traditional discrimination diagram managed only around 0.73 accuracy on the same task. The random forest’s advantage is likely rooted in its ensemble architecture: by averaging many decorrelated trees, it suppresses the noise from individual weak decision paths and captures nonlinear interactions among elements that no two- or three-axis diagram can represent.

Beyond raw accuracy, the study extracted geological insight from the models themselves. Feature-importance analysis of the random forest showed that both major-element differentiation and trace-element enrichment contribute jointly to the classification. This aligns with geochemical theory: major elements like silica reflect the degree and depth of partial melting, while trace elements such as niobium, thorium, and the rare earth elements record the nature of the mantle source and the imprint of subduction fluids. Machine learning, in effect, rediscovered the dual importance of source and process, but did so quantitatively and without human guidance on which elements should matter.

Perhaps the most provocative finding concerns the samples the models were least sure about. The researchers performed a probabilistic confidence analysis, examining cases where the classifier’s output probabilities were low or split between categories. Rather than representing random noise or data-entry errors, these low-confidence samples correspond to geochemically transitional or hybrid compositions, rocks whose chemistry sits between the canonical MORB, OIB, and IAB signatures. The authors interpret this as evidence that model uncertainty captures genuine tectonic information: ambiguous samples may record mixing of mantle sources, complex tectono-magmatic evolution, or settings where multiple processes overlap. In this reading, a machine learning model’s hesitation becomes a scientific signal, flagging samples that deserve closer scrutiny rather than discarding them as errors.

The geological coherence of the results reinforces this interpretation. The classification boundaries the algorithms learned map cleanly onto established mantle geochemistry. MORB samples cluster with depleted mantle signatures, reflecting the well-worn upper mantle beneath spreading ridges. OIB samples carry the enriched mantle fingerprints of deep plumes, with elevated incompatible element concentrations. IAB samples display the characteristic modifications wrought by subduction, including the involvement of crustal components and fluid-mobile element enrichment. The geology and the statistics tell the same story from independent directions.

The implications reach well beyond the ocean basins. Oceanic basalts are famously recycled into ophiolites, slices of ancient seafloor thrust onto continents, and distinguishing their original tectonic settings is essential for reconstructing supercontinent cycles and ancient plate configurations. Machine learning classifiers trained on modern samples can now be applied to these ancient rocks with far greater confidence than traditional diagrams allow. The approach also demonstrates a broader principle: by pairing the enormous public geochemical databases accumulated over decades with interpretable machine learning workflows, Earth scientists can turn archived measurements into quantitative tools that outperform the heuristic methods of the pre-digital era.

The study builds on a rapidly growing body of work applying machine learning to petrology, from neural networks that classify rocks from thin-section images to sparse-modeling approaches for tectono-magmatic discrimination. What distinguishes this contribution is its combination of a large curated training set, systematic comparison of multiple algorithms, honest evaluation on independent test data, and a deliberate effort to interpret not only the correct classifications but also the uncertainties. As the authors note, the integration of extensive geochemical databases with interpretable machine learning facilitates genuinely quantitative discrimination of oceanic basalt, a small but meaningful step toward making the Earth’s chemical archive fully searchable by algorithm, with geologists still supplying the questions.