Main

Oceanic anoxic events (OAEs) are transient perturbations to the global carbon cycle, during which large regions of the oceans are depleted in dissolved oxygen1,2. The resulting euxinic (anoxic and sulfidic) waters are highly toxic, leading to biological turnover events and, in extreme cases, mass extinctions of marine biota2. For example, the Toarcian OAE (T-OAE) involved an ~70% reduction in the diversity of some benthic fauna3. It is widely accepted that OAEs are ultimately linked to global warming2, which intensifies the hydrological cycle4 and chemical weathering on the continents5,6. These conditions lead to increased primary production and organic matter burial in the oceans7, documented in the geological record as organic-rich black shales1 (Fig. 1). These processes, in turn, cause a substantial reduction in deep-water O2 concentration ([O2]).

Fig. 1: Global distribution of OAE sedimentary deposits and plate boundary features.
figure 1

a,b, Plate-tectonic reconstructions (Methods) showing the main palaeogeographic features, MORs and exposed large igneous provinces, as well as the approximate distribution of OAE-related sequences in the Toarcian OAE at about 183–182 Ma (with OAE sites from ref. 49 and references therein) (a) and Turonian at about 90 Ma (with OAE sites from ref. 50 and references therein) (b). Note that shallow seas include epicontinental seaways, including the Western Interior Seaway of North America. CIE, carbon isotope excursion; HALIP, High Arctic Large Igneous Province. Figure created with GPlates26 (https://www.gplates.org/).

Full size image

Other Phanerozoic OAEs include the Steptoean Positive Carbon Isotope Excursion, the Hirnantian OAE, the Frasnian–Famennian and the Permian/Triassic boundary event4. However, the Mesozoic brought multiple OAEs in close succession (Figs. 1 and 2), with a mean duration of 1.4 Myr and a recurrence interval of ~7 Myr (Extended Data Table 1). Earth system models implicate a role for continental configuration in promoting deoxygenation of Mesozoic ocean shelves8. However, the interconnected nature of coeval environmental processes, including carbon outgassing from volcanogenic and/or methanogenic sources, global warming, increases in thermohaline circulation and changes in nutrient cycling, leaves the first-order drivers of OAEs open to debate.

Fig. 2: Tectonic, magmatic and oceanic chemical changes during the Mesozoic.
figure 2

a, Seafloor production rates (from ref. 6) and total length of MORs and transform faults51, and major phases of ocean basin formation (onset labelled). OAEs are shown as horizontal beige bands (note that the Kimmeridge is not strictly an OAE, and the dashed line for the Late-Hauterivian Faraoni OAE (F-OAE) is from ref. 52); b, LIPs (those with areas of >200 × 103 km2)11; also shown is the revised age range for the Ontong–Java Plateau53; c, δ13Corganic from organic-rich sediments54. d, δ13Ccarb (inorganic carbon) from marine carbonates55. e, The 187Os/188Osi signatures of sediments (data sources provided in Extended Data Table 2). f, Seawater 87Sr/86Sr ratios from marine carbonates56; these ratios gradually increased during the Mesozoic era in step with Gondwana fragmentation (see h). g, Continental fragmentation index calculated using continental perimeter/area from plate reconstructions57 (greater values signify a higher degree of fragmentation, and vice versa). h, Simplified plate reconstructions showing Gondwana breakup (orthographic projection) on the basis of the Ocean Drilling Stratigraphic Network magnetic reference frame (https://www.odsn.de/). Afr–Antarc, Africa–Antarctica; Afr–Mad, Africa–Madagascar; Aus–Mol, Australia–Molucca.

Source data

Full size image

OAEs and large igneous provinces

OAE initiation has been attributed to intensified carbon dioxide (CO2) emissions from large igneous provinces (LIPs)9,10 (Fig. 1), many of which coincide with known anoxic events (Fig. 2b). Uncertainties in LIP ages, typically ±1 Myr (similar to mean OAE duration), and intervals of activity up to 40 Myr (Fig. 2b)11 pose challenges for establishing conclusive linkages. A compounding factor is that LIPs are typically associated with continental rifted margins where other geological processes co-occur12. There is no doubt, however, that the high frequency of LIPs (Fig. 2b) contributed to the prolonged greenhouse conditions of the late Mesozoic era that promoted intense chemical weathering. The question is whether LIP emplacement initiates an instantaneous weathering response. Global chemical weathering intensity over the past 400 Myr is only weakly correlated with both the total eruptive area and LIP area within the tropics6, implicating other factors. However, the influence of LIPs peaks 35–50 Myr after emplacement6, perhaps indicating a delayed weathering response associated with passive margin rejuvenation13.

The Mesozoic era brought substantial changes in global ocean chemistry, including several carbon isotope excursions (Fig. 2c–f). These changes do not provide an unambiguous record of the relative magnitude of weathering inputs as geochemical proxies record contributions from multiple, simultaneous geological processes6. For example, the gradual increase in seawater 87Sr/86Sr (Fig. 2f) reflects increased weathering related to arc volcanism as well as continental breakup6 (Fig. 2h). Further, the long oceanic residence time of Sr (~5 Myr) limits its utility in identifying OAE drivers. While 187Os/188Osinitial (Osi) signatures offer more promise, a substantial osmium input from continental weathering sources could plausibly mask a volcanic signal (and vice versa), even if volcanism was the first-order driver of global anoxia.

Global tectonic reorganization of the Mesozoic

We propose that processes associated with continental breakup plausibly drove many Mesozoic OAEs (Fig. 2f). Breakup and mid-ocean-ridge assembly reached peak intensity during the major interval of OAEs (Fig. 2a,g,h). The ‘rift-to-drift’ transition and associated uplift of continental margins generates extensive volcanic terrains, which, in turn, lead to increased weathering of mafic lithologies. Volcanic rocks such as basalts are enriched in phosphorus (P)14,15, a biolimiting nutrient that regulates burial of organic matter over geological timescales16. Enhanced weathering of such rocks supplies nutrients to oceans, influencing bioproductivity17,18. For example, excess delivery of bioavailable P can lead to eutrophic conditions that deplete water column [O2], causing widespread anoxia or euxinia19. Such conditions can be sustained over million-year timescales via further recycling of P from sediments overlain by anoxic or euxinic waters, as indicated by observations20,21 and models19,22 of the Cenomanian–Turonian OAE (OAE 2). Like the weathering of mafic rocks on land15, volcanism during mid-ocean-ridge formation contributes essential nutrients, including P, directly to the global ocean through submarine weathering14.

Incipient MORs

During the Mesozoic, the breakup of Gondwana (Fig. 2f–g) involved a protracted, ~120-Myr-long episode of mid-ocean-ridge assembly on a global scale (Figs. 1, 2a and 3). In addition to fragmenting the continents and increasing global weatherability6, this process generated narrow ocean basins (Figs. 1 and 3a and Extended Data Fig. 1) where restricted water circulation can cause limited ventilation of bottom waters, locally encouraging the expansion of anoxia and euxinia. Although enhanced weathering can increase organic carbon burial, removing atmospheric CO2 and causing global cooling, this process may be counteracted by the expansion of the global ridge system—increasing net carbon outgassing into the oceanic reservoir23, contributing to increased dissolved inorganic carbon and driving ocean CO2 degassing on geologically short timescales24.

Fig. 3: Scale of incipient MORs globally during the Mesozoic.
figure 3

a, Plate reconstructions for the Central Atlantic, Africa–Antarctica and South Atlantic incipient MOR systems, with synthetic isochrons showing the time (in Myr) relative to the reconstruction time (labelled; absolute ages in Ma are provided in brackets); see Extended Data Fig. 1 for all active Mesozoic ridge systems. b,c, Total ridge length (b) and mean full spreading rates for ridges (c) based on ref. 58; values shown in b and c are means for each ridge system over 30 Myr (Methods and Supplementary Data 1). d, Summary timeline of shallow ridge volcanism; OAEs are shown as vertical grey bands, and the timing of precursory continental rifting33 is depicted with horizontal bars. e, Summed magmatic productivity of shallow ridges from 200 to 60 Ma incorporating age uncertainty (Methods), assuming water depths of <2 km for 30 Myr (right side); longer-term context back to 350 Ma (left side). f, Modelled combined ridge flux of P for all ridges active during this period (Extended Data Figs. 1 and 3), calculated in 10,000 simulations using ridge productivity data shown in e. Simulations account for uncertainty in the time of ridge initiation (Methods). P–T, Permo–Triassic anoxic event (linked to Siberian traps LIP); PETM, Palaeocene–Eocene Thermal Maximum (linked to North Atlantic LIP31), thought to represent an OAE2.

Source data

Full size image

Extensional plate boundaries are a key source of nutrients to the oceans during breakup14. When continental rift basins submerge below sea level, it results in extensive magma–seawater interactions and accelerated chemical weathering of volcanic regolith. In such transitional environments, referred to here as incipient mid-ocean ridges (MORs), enhanced volcanism driven by magma–seawater interactions, coupled with low hydrostatic pressures, also promotes the explosive eruption of basalt25. To explore the impact of MOR assembly on ocean chemistry, we first investigate the spatial extent of mature continental rifts that transitioned to nascent ridges within relatively shallow ocean basins, defined here as <1,500 m deep (Extended Data Fig. 1 and Methods). We identify nine incipient ridge segments using the plate-tectonic reconstruction software GPlates26 (https://www.gplates.org/). We find ~26,000 km of MOR initiated during the Jurassic–Cretaceous, 85% of which was simultaneously active from 160 to 120 million years ago (Ma) (Fig. 3b–d). Analysing seafloor production rates along these ridges, we find that the main incipient MORs overlap in timing with the main OAE interval (Fig. 3d). In northern latitudes, this includes a prolonged phase of oceanic anoxia within the Kimmeridgian, which, while not strictly an OAE, is considered one of the most prominent organic carbon burial episodes of the Phanerozoic27. During OAE 2, δ13C excursions—signalling changes in both local and global carbon cycles—are attributed to basalt–seawater interactions20, consistent with coeval sharp declines in Osi in marine sediments28 (Fig. 2e). Here expanded deoxygenation preceded massive organic carbon burial, reflecting dynamic changes in oceanic redox conditions29.

Within incipient MOR environments, fragmental, glass-rich basalts undergo rapid alteration, releasing a substantial fraction of elements, including phosphorus, into seawater14 (Extended Data Fig. 2). Further, incipient MORs coincide with the rift-to-drift phase when decompression melting and upwelling of asthenospheric melts tend to be highest30 and are typically characterized by an order of magnitude increase in magmatic productivity31. This phase of seafloor volcanism can plausibly transform large domains of oceanic crust from minor sinks32 into net sources14 of phosphate (Fig. 2d–f). Hyaloclastite breccias and other pyroclastic rocks are commonly formed during the transition from continental rifting to seafloor spreading and have been linked to early explosive magma–water interactions during submergence of the spreading axis25. For example, during the formation of the Central Atlantic Ridge33 (Fig. 3a)—which unzipped to create the Hispanic Corridor (Fig. 1a)—a massive wedge of fragmental basaltic rocks was emplaced along the Atlantic margins of North America. With a volume of ~3 × 106 km3, this feature is one of Earth’s largest igneous provinces30.

We use seafloor production data to estimate P release from basalts emplaced along global ridges (Fig. 3). We use modelled lengths and spreading rates for each ridge to calculate seafloor ridge productivity over time (Fig. 3b–c). We simulate a linear decay in fragmentation and phosphorus release during seafloor subsidence, accounting for reduced weathering intensity as the ridge progressively deepens (Methods). This analysis indicates that P release (Fig. 3f and Extended Data Fig. 3) potentially exceeded 3 × 1010 mol yr−1, equivalent to over 75% of the current total reactive P flux from rivers to the oceans (~3.9 × 1010 mol yr−1) (ref. 34).

Subaerial basalt uplift and weathering

Subaerial basalt weathering can deliver a substantial flux of bioavailable P to the ocean15, boosting primary production and organic carbon burial and driving global cooling35. Box models of oceanic phosphorus, carbon and oxygen cycles have demonstrated that increased P delivery from continents is sufficient to drive OAEs19. Alongside our ridge-derived P flux, we propose an additional riverine supply from a well-established phase of tectonic uplift and exhumation across Mesozoic southern Africa13, which caused rapid denudation of thick subaerial flood basalts36,37,38. Previous thermochronological studies reveal multiple phases of abrupt cooling of rocks that correspond to erosion of approximately 2-km-thick basalts of Karoo-age (~183 Ma) (refs. 36,37,39) (Methods and Fig. 4a–c), long (>30 Myr) after their eruption. A comparison between denudation rates and offshore sediment accumulation36 indicates that three times more material was eroded than deposited in offshore basins at this time37. This deficit has been attributed to intense chemical denudation37, consistent with humid subtropical climates regionally at this time40. This interpretation is supported by two geological observations: (1) coeval deltas dominated by clays derived from weathering of mafic lithologies37; (2) widespread chemogenic carbonates in offshore basins, locally >500 m thick and covering ~330,000 km2, purportedly derived from Cretaceous basalt alteration41.

Fig. 4: Chemical denudation of continental basalts in the Mesozoic.
figure 4

a, Distribution of Drakensberg basalts and their feeder intrusions (dark blue) within the Karoo Basin of South Africa. b, Schematic cross section reconstructions at 120 Ma (top) and ~110 Ma (bottom) (line of section in a) based on apatite (U–Th)/He thermochronometry39 and xenoliths in kimberlite pipes. c, Estimated cooling rate of rocks across South Africa from thermochronologic models13 (Methods) related to erosional exhumation between 160 and 80 Ma (hence, age interval shown differs from Fig. 3), showing cooling trends associated with multiple regional events (Methods); the timing of OAEs is shown for context (the earliest event shown, the Kimmeridge (adopting the age of that stage from the International Commision on Stratigraphy, v2023/09), is not strictly an OAE but involved prolonged phases of oceanic anoxia). d, Modelled annual flux of P from basalt denudation shown in a, utilizing modelled cooling trends from thermochronology for individual localities across southern Africa to drive denudation rate (Methods). Panels a and b adapted with permission from ref. 39, John Wiley & Sons.

Source data

Full size image

We next estimate the phosphorus release from weathering related to tectonic uplift in southern Africa (Fig. 4b,c and Methods). Here major exhumation occurred between 120 and 85 Ma, with peaks at about 120, 112, 104 and 95 Ma (Fig. 4c), and denudation rates up to 125–175 m Myr−1 have been reported36. At this time, basalts encompassed the wider Karoo Basin—as evidenced by thermochronological constraints37,39, basalt fragments preserved in deeply eroded kimberlite pipes and the present-day footprint of feeder intrusions (Methods). We develop a simple model using existing thermochronological ‘best fit’ models to drive denudation rates (Fig. 4c), sampling observed P2O5 concentrations from a global compilation of Karoo-aged basalts (GeoROC Database; Methods). Our models estimate that P fluxes from the chemical denudation of the basalts potentially reach 3 × 1010 mol yr−1 (Fig. 4d)—a similar magnitude to the present-day dissolved inorganic P input to the ocean34. Previous studies have shown that P accumulation rates rise and reach a peak coinciding with the onset of OAE 2 (ref. 21). Our model suggests this could be due at least in part to increased delivery of phosphate from continents, giving rise to increased primary productivity, organic carbon burial and black shale formation21.

While it is important to recognize the influence of other palaeogeographic features, such as the Central Pangaean Mountains (Fig. 1a), on continental weathering, constructing a fully integrated Earth system model is beyond the scope of our study. Consequently, our estimated phosphorus fluxes for this comparatively well-documented phase of erosion (Fig. 4) should be regarded as conservative minimum estimates. Considering age uncertainties within terrestrial and marine records (typically ±1–2 Myr), we observe good overall correspondence between the timings of peak weathering events and OAE occurrences (Fig. 4c,d). This prompts the question of whether this relationship is coincidental or causal.

Biogeochemical model

To interrogate whether basalt weathering in submarine and subaerial environments could have driven anoxia, we test the influence of phosphorus release events on oceanic [O2] and examine how their timings relate to OAEs in the geological record (Fig. 2). We incorporate both submarine (Fig. 3 and Extended Data Figs. 3 and 4) and terrestrial (Fig. 4) basalt weathering fluxes (Supplementary Data 2 and 3) into a global biogeochemical model of carbon–oxygen–phosphorus cycles developed by ref. 42 and adapted by ref. 43 (Extended Data Fig. 5). The model divides the ocean into proximal and distal shelf zones, a surface open ocean and a deep ocean interior and explicitly represents primary production and remineralization, as well as the decreasing retention of P in sediments under anoxic conditions21. The ridge flux of P is injected into the model’s deep ocean box, while the riverine weathering flux is delivered into the proximal shelf zone box. Following ref. 19, we amend global ocean circulation to account for the warmer, more stagnant ocean of the later Mesozoic. We reduce open ocean vertical mixing to 70% relative to present day, which falls within sensitivity estimates of previous studies19. All other background forcings follow the long-term Earth evolution scenario from ref. 43 (Methods).

Figure 5 shows model runs through the Jurassic and Cretaceous incorporating both ridge (Fig. 5a and Extended Data Fig. 6) and riverine (Fig. 5b) P fluxes. Each ridge initiation and basalt uplift/weathering event is considered independently (Methods, Fig. 5 and Extended Data Fig. 7) because the biogeochemical model does not yet fully incorporate feedbacks for atmospheric oxygen, for example, terrestrial biosphere and wildfires44. Because organic carbon removal means liberation of O2, later events would be applied in a modified high-oxygen state, which is not consistent with the geological record of broadly stable atmospheric oxygen levels across the Jurassic and Cretaceous45. When viewing the model output in these terms, calculated P fluxes and input rates are sufficient to trigger substantial ocean deoxygenation (Extended Data Fig. 7). For T-OAE, our model predicts a prolonged phase of anoxia (Fig. 5a) that conforms with global-scale deoxygenation indicated by marine redox proxies, extending well beyond this event3,46. MOR initiation, along with the climate-altering impacts of Karoo–Ferrar volcanism3,46 (Fig. 2b), can jointly reconcile the long-term unradiogenic Osi signal bracketing the T-OAE.

Fig. 5: Modelled ocean deoxygenation through time and observed timing of OAEs.
figure 5

a,b, Mean, 5th–95th percentile and minimum–maximum estimates for dissolved [O2] in the ocean interior over time from biogeochemical model simulations (see Methods for details). We input multiple (1,000) simulated phosphorus release events for shallow MOR systems (a) and riverine (continental weathering) (b) sources tied to exhumation in southern Africa during and after Gondwana breakup (Fig. 4); OAEs are shown as vertical grey bands (labelled); upper plots show the P fluxes as discrete events resulting from both processes (see Figs. 3f and 4b for cumulative fluxes). c, To evaluate the power of the (normalized) minimum value of [O2] (see Extended Data Fig. 8) as a predictor of OAEs, we plot an ROC curve, shown. The AUC of 0.705 shows that the [O2] minima in our models offer reasonably good predictive performance for OAEs in the geological record. The table summarizes true negative rate (TNR) and true positive rate (TPR) for the predictor using two alternative thresholds (Methods).

Source data

Full size image

The overall correspondence between predicted deoxygenation events and OAEs (Extended Data Table 1) appears moderately good. Modelled drops in [O2] broadly coincide with 9 out of 11 confirmed OAEs (Fig. 5). Notably, OAEs 1c and 1d do not overlap with either ridge or riverine deoxygenation, occurring approximately 2 Myr before the modelled peak related to riverine inputs (Fig. 5b). Further, OAE 1b occurs ~3 Myr before the modelled peak in deoxygenation. However, it is important to consider the age uncertainties in uplift/weathering events and OAE ages (typically ±1–2 Myr). OAE 2 coincides with an apparently smaller-scale weathering event (Fig. 5). At the Jurassic/Cretaceous boundary, we predict a relatively minor deoxygenation event linked to ridge volcanism, despite no known OAE in this interval (Fig. 5a). To quantify model performance, we apply a simple classifier derived from estimated [O2] minima to predict OAE occurrence. We identify the minimum value of [O2] over all individual ridge and weathering model runs (Extended Data Fig. 8) and normalize this to give a value on the interval [0, 1], which can be used as a predictor of OAE occurrence (a binary outcome). Performance is measured using the receiver operating characteristic (ROC) curve (Methods), where the area under the curve (AUC) of 0.705 (95% confidence intervals: 0.660–0.750) signifies a good but imperfect overall predictive value (Extended Data Fig. 8) for OAE timing, outperforming random chance (Fig. 5c).

One feature of the OAE geochemical record that is difficult to reconcile with any model for their origin is the presence of both negative and positive excursions in δ13Ccarb (inorganic carbon) values20 (Fig. 1c,d). In general, increased productivity and eutrophication are expected to yield a positive δ13Ccarb signature (for example, OAE 2) as predicted by our model. Hence, the presence of negative δ13C requires a source of isotopically light carbon. Individual volcanic events may also release light δ13C if they trigger processes such as permafrost melting and/or thermal metamorphism of mudrocks or coal47, but this does not negate the role of P release through volcanic weathering in stimulating productivity and causing anoxia/euxinia.

Extent of seafloor anoxia

Our results suggest that most OAEs could be explained by an interplay between enhanced weathering of mafic rocks in submarine and subaerial environments (Fig. 5). Model uncertainty indicates that deoxygenation events yield mean decreases to as little as ~110 μmol kg-1 [O2] averaged across the ocean interior, representing an ~30% decrease in average marine [O2]. As marine [O2] is difficult to infer for the deep past, most geological estimates of OAE intensity instead focus on the fraction of seafloor overlain by anoxic water. This is also uncertain but appears to be between 1% and 10% of the seafloor for most events4. A simple estimation considers that present-day anthropogenic deoxygenation has reduced mean [O2] by about 2%, leading to an 85% expansion of anoxic seafloor48. Thus, an ~30% reduction in [O2] is expected to increase seafloor anoxia from the present 0.3% of the ocean floor to about 4%, which is broadly consistent with the magnitude of OAEs in the geological record4,46. Modelled probability distributions (Fig. 5a,b) show that basalt weathering can drive much higher degrees of deoxygenation and, in turn, more severe seafloor anoxia (Extended Data Fig. 6).

The proposed weathering mechanisms do not necessarily contradict models for OAE triggering that invoke episodic massive volcanism. Our model quantifies a complimentary source of P that could theoretically drive organic carbon accumulation in marine basins and substantial deoxygenation. This does not preclude the involvement of additional disturbances in the global carbon cycle; rather, it may operate alongside them, if not directly driving OAEs, making oceans more susceptible to anoxia triggering.

To conclude, we hypothesize that continental breakup resulted in elevated background phosphorus levels during the main interval of Mesozoic OAEs (Figs. 2 and 3) sufficiently potent to drive both prolonged and abrupt phases of deoxygenation (Fig. 5). Specifically, multiple pulses of continental basalt weathering linked to Gondwana breakup (Fig. 4) coincide closely in time and magnitude with OAEs (Fig. 5). Both processes plausibly combined to deliver pulsed fluxes of nutrients to the oceans, driving high rates of organic carbon burial and recurrent anoxic events. The Upper Cretaceous decline of this tectono-magmatic activity may help reconcile the frequency and intensity of OAEs in Mesozoic oceans. Our results thus support a mechanistic link between solid Earth processes and anoxic events that caused profound biological turnover episodes1,2,3. The proposed feedbacks between continental reorganization, magmatism and anoxia warrant further investigation and carry implications for understanding the environmental consequences of other breakup events during Earth history.

Methods

Quantitative estimation of seafloor production rates

To calculate seafloor production rates, we used the plate model of ref. 58, with some corrections as outlined in ref. 59. This plate model comprises a digital repository of continents and plate boundaries and, through a series of Euler rotations, tracks how they have evolved through time. For the Mesozoic–Cenozoic, they are based predominantly on the seafloor spreading history preserved in ocean basins60.

In our analysis, we use the Python library pyGPlates (gplates.org/docs/pygplates/index.html) to extract all MORs where seafloor spreading occurred, and their associated kinematic data, from the plate model. This includes the length of spreading segments, the rate of spreading, start and end time of activity and the reconstructed position (the latitude and longitude of the MOR), among other metadata. Overall seafloor production rates—a measure of the area in kilometres squared of new seafloor formed along MORs per million years—are calculated by summing the product of spreading rate and spreading segment length for each individual length of the global ridge system. We then export the entire dataset to individual .csv files for each 1 Myr time step between 200 and 0 Ma.

Using the plate-tectonic reconstruction software GPlates26 (https://www.gplates.org/), we defined bounding polygons around incipient MOR systems that evolved from a continental rift (in a region of thick continental lithosphere; Extended Data Fig. 1). In our model, these polygons are active for the first 30 Myr of spreading, when the incipient ridge is mostly still relatively shallow and susceptible to enhanced weathering (Main). Nevertheless, we apply a linear decay in fragmentation and phosphorus release during seafloor subsidence (as discussed in the following), which accounts for reduced weathering intensity as the ridge progressively deepens in our model. Spatially, the polygons are defined around congruent geographic regions representing incipient ocean basins. These geographic regions do not always map well to present day; for example, the opening of the Atlantic Ocean is divided into Northern, Iberian, Central, Equatorial and Southern ridge segments. The oceanic ridge systems included in our analysis, with accepted start times in brackets, are as follows: Central Atlantic (180 Ma), Australia–Molucca (155 Ma), Africa–Madagascar (154 Ma), Africa–Antarctica (154 Ma), Iberian Atlantic (140 Ma), South Atlantic (132 Ma), India–Australia–Antarctica (131 Ma), Equatorial Atlantic (120 Ma) and India–Madagascar (87 Ma). These data are summarized in Fig. 3a–d and Extended Data Fig. 1 and are supplied in Supplementary Data 1.

The preceding analysis yields a series of polygons that spatially and temporally define the limits of incipient MORs and a series of datasets describing all active ridges at a specified time step. We use the ‘select’ function of the Generic Mapping Tools (https://www.generic-mapping-tools.org/) software, which intersects datasets (including .csv files with latitude/longitude coordinates) against a polygon(s) to extract data within the polygon at each time step per bounding box (at their active times). This gives us individual time series (multiple .csv files) of the ridge segments within each bounding box that we merge into individual profiles for each discrete ocean basin.

Mapping OAEs

Figure 1 plots the location of organic-rich facies and carbon isotope excursions of the two major OAEs, T-OAE and OAE 2. These locations were previously compiled by ref. 49 for T-OAE (building on the earlier compilation of ref. 61) and ref. 50 for OAE 2. The maps were produced using the open-source plate-tectonic reconstruction software GPlates, using the palaeogeographic regions of ref. 62 and the plate rotations of ref. 63. The distribution of large igneous provinces was plotted using the shapefiles of ref. 64.

Modelling shallow MOR phosphorus release

To estimate the flux of phosphorus into the ocean from fragmented basaltic materials emplaced along shallow incipient MORs (Extended Data Fig. 1), we adapted models from ref. 14, which probabilistically estimate elemental fluxes from the chemical weathering of seafloor basalts, accounting for uncertainties in parameters. Because seafloor production is reasonably well constrained for the Jurassic and Cretaceous58,65, we used seafloor production data (acquired using the preceding procedures) in our models. These data include estimates for seafloor production rates (km2 Myr−1), total ridge length (km) and mean spreading rate (km Myr−1) globally at 1 Myr intervals. The nine ridges considered in this analysis, listed in the preceding (Fig. 3a,d), are shown in Extended Data Fig. 1 (Supplementary Data 1). We define ‘shallow’ as water depths of 1,500 m or less. We also consider durations for ridge productivity of up to 30 Myr after ridge initiation, a period during which the ridge should be at its shallowest66,67, particularly in the presence of dynamic support. To account for age uncertainty associated with the timing of ridge formation, we apply a random offset to the start time (ts) of each ridge (Fig. 3e), sampling from a beta distribution scaled to the interval [0, 0.05 ts]. This approach simulates uncertainty that increases further back in time6, up to 5% of ts. We use beta parameters α = 1, β = 1.5 (skewed towards zero), which corresponds to a mean offset of 2%, and standard deviation of approximately 1.3%.

Our model estimates the volume of fragmental basaltic material over time for each rift/ridge segment (Main). We assume that the total thickness of this material formed and accessed by seawater infiltration will decrease over the duration considered in the model runs. This assumption is considered reasonable because (1) higher magmatic productivity is expected at the earliest stages during ridge initiation due to decompression melting and active upwelling of asthenosphere30,31,68, which typically coincides with (2) initially shallow water conditions that favour more intense hydrovolcanism (for example, hyaloclastite formation)25 and pyroclastic eruptions14; (3) when new seafloor subsides and is transported away from the ridge it cools and is expected to experience less vigorous convection by infiltrating seawater and hydrothermal flow processes69,70. Therefore, we simulate a simple linear decay in thickness of fragmental material over time, from 1,500 m to 100 m over the model run, values that are considered realistic on the basis of observations of seafloor hydrothermal systems14,69. We apply ranges of values for shallow and deep porosity, hyaloclastite abundance and palagonite fraction71 used in previous models14. We also used element loss estimates from ref. 14, based on ref. 71 (Extended Data Fig. 2), yielding the following uniform distributions: P2O5 loss: [0.002, 0.006]. In our calculations, we use a standard density of 2,390 kg m−3 for P2O5.

We performed 10,000 individual calculations to build an output distribution that captures the estimated uncertainty in model parameters and timing of ridge initiation. For each iteration, we step through each ridge in turn, sampling palagonite fraction, shallow/deep porosity, hyaloclastite fraction and element losses (specifically, in our case, P2O5). Then, for each ridge, we step through the active period at 1 Myr intervals. At each time step from t = 0 (when the ridge first becomes active), we use the ridge production values and alteration depth (using the linear decay model) from the input file. The absolute event timing is modified for each ridge at each iteration using a sampled offset (explained in the preceding), and the time series are finally interpolated to regular 1 Myr intervals before being combined with the results from all ridges and model iterations.

We next calculated hyaloclastite thickness as a function of alteration depth and sampled hyaloclastite fractions. By multiplying hyaloclastite thickness by ridge production, we then calculate the total hyaloclastite volume and hyaloclastite weight. The total weight of P2O5 released is the weight of hyaloclastite multiplied by the randomly sampled palagonite fraction and P2O5 loss (using the distributions given in the preceding). We then convert to moles P as a total flux into the ocean per 1 Myr.

This procedure gives 10,000 individual (offset) time series for P for each ridge. We performed linear interpolation of each (offset) time series to estimate the total P (mol) released from each ridge at regular 1 Myr time steps from 180 to 59 Ma (Extended Data Fig. 3). Then, for each individual iteration, we calculated the total mol P from all active ridges. This gives 10,000 variant time series for total estimated mol P. From this we can then calculate the mean, median and percentile estimates at each time step (5th, 10th, 90th and 95th percentile mol P). The resulting time series for mol P, and associated uncertainty, is shown in Fig. 3f. These data are supplied in Supplementary Data 2.

Modelling continental denudation

In addition to incipient MORs, we simulate the phosphorus contribution from a major phase of uplift and denudation that occurred across southern Africa from about 160 to 80 Ma (Main). This denudation involved basaltic lavas of the ~183 Ma Karoo Group72. The present-day Drakensberg Mountains in Lesotho are a relic of this thick basaltic sequence, which during the Cretaceous period extended much farther to the west39. We sought to quantitatively estimate the amount of P liberated during chemical denudation of these basalts in this interval. To do this, we turned to thermochronology, which provides constraints on the spatial and temporal extent and magnitude of crustal cooling and thus exhumation (uplift and denudation). These cooling history models36,37,39,73,74,75,76,77,78,79,80 cover a large area of southern Africa owing to the availability of well-dated kimberlite pipes.

We used the rate of change of temperature (cooling rates) from previous thermochronology studies for 28 individual localities spanning southern Africa, using best-fit thermal histories defined in those studies13. These are temporally irregularly spaced observations/measurements, so we interpolate to give regular time series with a 0.1 Myr interval (a linear interpolation using the R approx() function). We extend all time series to span the period 160 to 55 Ma using the first and last observed values to pad the series as required.

For each locality we use the R function rollapply() to calculate the change in temperature T over a 1.9 Myr symmetric (rectangular) moving window:

$${{\Delta }}T/{{\Delta }}{t}_{1.9{\mathrm{Myr}}(t)}=\frac{(T(t-0.9)-T(t+0.9))}{1.9}$$

This is the estimated temperature drop occurring at time t over the window t ± 0.9 Myr. Note that the total width is an odd number as we require a symmetric window. At each 0.1 Myr time step, t (from 160 to 55 Ma), we then find the maximum value of ΔT/Δt1.9Myr over all the localities and save this as a separate time series (max(f(t)), where f(t) = maximum temperature drop over all localities at a given time step, t). This is then used to estimate the denudation rate and quantitatively compute how this varies over time.

We assume a minimum denudation rate Dmin = 20 m Myr−1, and maximum rate Dmax = 175 m Myr−1, on the basis of previously reported denudation rates in this region spanning this interval and underpinned by thermochronology and offshore sediment accumulation histories36. We make the simplifying assumption that the denudation rate scales linearly with the rate of temperature drop—the estimated maximum ΔT at each time step. This means we have a denudation rate D = 20 m Myr−1 where ΔT is at its minimum (minΔT = min(f) over the period of interest) and a denudation rate D = 175 m Myr−1 where ΔT is at its maximum (maxΔT = max(f)). The scaling factor SD for denudation is then computed as follows:

$${\mathrm{Dscale}}(t)=\frac{(\,f(t)-{\mathrm{mi{n}}}_{{{\Delta }}T})}{({\mathrm{ma{x}}}_{{{\Delta }}T}-{\mathrm{mi{n}}}_{{{\Delta }}T})}$$
$${\mathrm{Denudation}},D={D}_{{\mathrm{min}}}+({S}_{D}\times (D_{\mathrm{max}}-D_{\mathrm{min}}))$$
$${\mathrm{Erode{d}}}_{{\mathrm{vol}}(t)}={\mathrm{Basal{t}}}_{{\mathrm{area}}}\times ({\mathrm{Denudation}}(t)/1,000)$$

The variable Basalt_area denotes the spatial extent of basalts in the Drakensberg Group. This area is estimated as 600,000 km2 on the basis of thermochronology models39,73, fragments of basalt preserved in deeply eroded kimberlite pipes81 and the present-day spatial footprint of intrusions that served as feeders to flood basalt eruptions72 (Fig. 4a). We convert denudation rate from m Myr−1 to km Myr−1 to give the total volume in km3 Myr−1. We can then estimate how much phosphorus this potentially involves.

To estimate P2O5 contents of the weathered basalts, we sample from a large database, obtained from the GeoRoc repository (https://georoc.eu/), of measured P2O5 contents from basalts of the Karoo and age-equivalent Ferrar Provinces. We used this dataset as an empirical distribution for P2O5 in our models instead of fitting an arbitrary distribution. We performed 10,000 iterations to build an output distribution for output P (Mol) over time. The P2O5 dataset has a total of 2,495 observations (latitude, longitude and P2O5 content as a percentage, ranging from >0 to 5.39%), and we exclude a small number of cases (36) with zero P2O5 contents. The estimated phosphorus fluxes from each of the nine continental weathering ‘events’ considered are supplied in Supplementary Data 3.

Biogeochemical modelling of the effects of P inputs

To test the effects of P input into the ocean through incipient seafloor ridges and continental basaltic weathering, we use the five-box ocean–atmosphere model of ref. 43. The model tracks the long-term evolution of the carbon, phosphorus and oxygen cycles over Earth history and was extended from a model originally developed to investigate Mesozoic OAEs42. In this study, we use the standard run of the model but impose phosphorus input events on the basis of our data compilation for ocean ridges (Fig. 3) and continental denudation events (Fig. 4). We include the sluggish open ocean upwelling based on previous uses of this model to investigate OAEs42. P input from ridges is delivered into the ‘Deep ocean’ box as soluble reactive phosphorus. P input from weathering is delivered into the ‘proximal zone’ box of the continental margin, again as soluble reactive phosphorus (Extended Data Fig. 5). We track the model response to these input forcings by plotting the deep ocean [O2] concentration (Fig. 5a,b, Extended Data Fig. 7 and Supplementary Data 4). Although we produce estimates of the combined P inputs over the Mesozoic, we test each input event in the model individually. This is because the model lacks powerful feedbacks on atmospheric oxygen levels associated with the land biosphere, and thus combined long-term increases in P inputs can raise atmospheric O2 to high levels that are no longer consistent with a Mesozoic background state.

Evaluating the predictive performance of biogeochemical models

Here we outline the methodology used to assess the predictive power of the minimum value of [O2] in our biogeochemical models (Fig. 5) as an indicator for the timing of OAEs in the geological record (Extended Data Table 1). To generate a single predictive measure from separate model runs, we first computed the mean output for the weathering and individual ridge model simulations (Fig. 5; note each ridge is simulated separately). At each time step, we identified the overall minimum value of [O2], selecting the lowest value from either the ridge model (Fig. 5a) or the riverine/weathering model (Fig. 5b). This allows us to derive a single measure from multiple, separate model runs (Extended Data Fig. 8a), mitigating the need to use arbitrary thresholds to define ‘lows’ in [O2] levels.

We then normalized this derived time series (Extended Data Fig. 8a) to the interval [0,1] and inverted the values (1 – normalized value). This yields a simple predictor for OAE events (a binary outcome), allowing us to test the hypothesis that values closer to 1 (corresponding to the [O2] minima) indicate a higher probability of OAE occurrence. This normalized and inverted time series is shown in Extended Data Fig. 8b.

To evaluate the predictive performance of this model, we plotted an ROC curve (Fig. 5c), which shows the trade-off between sensitivity (the true positive rate—the probability of predicting an OAE event given an actual OAE event) and specificity (the true negative rate, correctly predicting the absence of an OAE event) for a range of classification thresholds. The AUC provides a measure of the forecast performance: a random, no-skill forecast would follow the diagonal line with an AUC of 0.5, and a perfect classifier would have an AUC of 1 (perfect sensitivity and specificity). The 95% confidence intervals are computed from 2,000 stratified bootstrap samples using the R package, pROC82 (CRAN.R-project.org/package=pROC). The ROC plot (Fig. 5c) shows that the simple classifier (described in the preceding) derived from modelled [O2] minima (Extended Data Fig. 8) is a reasonably good predictor of OAEs in the geological record, surpassing random chance.

Two examples of potential diagnostic thresholds are given in Fig. 5c (corresponding to Extended Data Fig. 8). Youden’s J statistic83 (red text/line) defines the optimal cut-off as the threshold value that maximizes the distance to the diagonal. ‘Closest to top left’ (blue) minimizes the distance from the top left corner (the point corresponding to perfect sensitivity and specificity). Classifier values over a given threshold correspond to a prediction of an OAE event occurring, and values below the threshold predict no OAE.

As shown in the right-hand panel of Fig. 5c, using the threshold 0.19 gives a true positive rate (TPR) of 0.71; this is higher than using the Youden method (TPR 0.625), but at a cost of a lower true negative rate (TNR 0.70 versus 0.79 for the Youden method). In practice, the relative cost of a false alarm or miss would be evaluated to determine the optimal threshold.