Climate scientists just learned to weigh time itself — and the Little Ice Age ocean got colder

A thin column of ice drilled from deep beneath West Antarctica holds the memory of centuries of surface temperature. When you pull a borehole profile out of the ground, you're not reading a single year — you're reading a blurred, diffused echo of decades and centuries of warmth and cold, compressed by the physics of heat moving slowly through rock and ice. For a long time, climate scientists have struggled to put this long-memory record on the same playing field as an annual tree ring, which records a single growing season in a single year. They measure different things, on different timescales, and no one could agree on how to weigh them together.

Now a team led by Zilu Meng at the University of Washington has built a framework that finally treats every paleoclimate archive for what it actually is — not as an instantaneous snapshot, but as a witness with its own temporal memory. Called the Last Millennium Reanalysis 4D-Var (LMR4D-Var), the framework reconstructs climate trajectories from 500 BCE to 2000 CE by simultaneously assimilating annual tree rings, decadal-to-millennial sediment and pollen records, and long-memory borehole temperature profiles. It does this, the authors report, while achieving the highest skill score against instrumental observations of any paleoclimate data-assimilation reconstruction to date — and it changes one of the most contested numbers in climate history.

That contested number is the ocean. When borehole constraints enter the reconstruction, the global-mean ocean during the Little Ice Age (roughly 1400–1700 CE) gets measurably cooler — by about $0.15\times 10^8$ J m⁻² in the shallow 0–300-m layer and $0.20\times 10^8$ J m⁻² in the deep 300–2000-m layer (Meng et al., 2026). This is not an arbitrary tweak: it emerges naturally from the physics of the coupled system, transmitted from long-memory surface constraints into subsurface heat storage through the model's dynamics, with no explicit link between boreholes and ocean heat ever programmed in.

The Science

The problem that LMR4D-Var solves is older than any single study. Paleoclimate archives are gloriously heterogeneous. Tree rings and corals resolve individual seasons; sediment and pollen records average over decades or centuries; borehole temperature profiles integrate centuries of surface history through thermal diffusion. A reconstruction framework that treats all of these as instantaneous observations risks either damping low-frequency climate signals or distorting high-frequency ones (Meng et al., 2026).

The dominant approach in recent paleoclimate data assimilation has been sequential filtering — a method that processes observations one time step at a time. That's appropriate for real-time weather prediction, when you genuinely can't see into the future. But it's fundamentally one-sided for reconstructing the past. Once a filter has passed a given year, any observations dated later are locked out of revising that earlier state. A borehole profile contains information about centuries of history, yet a filter treats it as an update at a single moment, discarding most of its temporal information.

LMR4D-Var instead formulates reconstruction as a trajectory-smoothing problem using a technique from numerical weather prediction called four-dimensional variational data assimilation — 4D-Var, for short. Rather than analyzing each observation time independently, 4D-Var finds the single climate trajectory that best fits observations scattered across an entire assimilation window at once. Because the whole trajectory is optimized together, an observation dated 1750 can revise a state from 1450, through the modeled temporal relationships that connect them.

The framework uses "weak-constraint" 4D-Var, which is an important technical choice. Strong-constraint 4D-Var assumes the model is perfect and lets only the initial condition vary. Weak-constraint 4D-Var instead estimates a sequence of model-error increments alongside the initial state, acknowledging that the reduced-order climate emulator — a simplified statistical stand-in for a full climate model — is imperfect. The optimization balances three competing penalties: the initial-state constraint, the model-error penalty, and the mismatch between model and proxies. Their relative weights are set by error covariances, so records with different error characteristics can be weighed within a common optimization.

The state vector tracks seasonal climate through a coupled system: surface temperature, sea-surface temperature, sea ice, and two layers of ocean heat content (0–300 m and 300–2000 m). The reconstruction is repeated five times, once for each of five climate-model-specific emulators, to assess model dependence.

Each proxy archive connects to the candidate trajectory through an observation operator — a mathematical function that represents what that archive actually records and over what interval. This is the conceptual heart of the paper: the archives retain their distinct temporal support instead of being forced to behave as instantaneous equivalents. Because the implementation is differentiable, the authors note it opens a practical route to adding nonlinear, time-averaged, or history-dependent proxy system models without needing a separate adjoint model for each archive.

What They Found

The authors validated the framework three ways: a controlled pseudo-proxy experiment, verification against withheld observations, and comparison with independent instrumental and reconstruction products. The team was careful to distinguish validating the assimilation framework from evaluating the resulting reconstruction.

In the pseudo-proxy experiment, one member of the CESM Last Millennium Ensemble was treated as a known "ground truth," its proxy values generated from that target, and the entire model family excluded from the training pool. The framework had to recover a known coupled climate trajectory — including ocean heat content that was never directly assimilated — from sparse, surface-sensitive, proxy-like information. It succeeded: correlation with the target was r = 0.71 for global-mean temperature, r = 0.69 for shallow OHC, and r = 0.71 for deep OHC over 1000–2000 CE, recovering several major low-frequency departures including volcanic responses (Meng et al., 2026). The point was never to prove the reconstruction is right, but to prove the machinery can recover known signals from information as sparse and indirect as real proxies provide.

Against instrumental-era observations, the framework performs strongly. The P2k_BH_T12k reconstruction (all proxy classes together) agrees with Berkeley Earth surface temperature at r = 0.95 with a coefficient of efficiency of 0.89. Strikingly, after linearly detrending both series, agreement remains high (r = 0.84, CE = 0.70), showing that the match is not just a shared long-term warming trend (Meng et al., 2026).

Global mean temperature skill vs. earlier reconstructions

Label	Value
LMR4D-Var (5 models)	0.95
LMR4D-Var (CCSM4)	0.94
LMR Seasonal (30)	0.93
LMR v2.1 (58)	0.93
LMR Online (48)	0.9
PHYDA (56)	0.9

Source: Meng et al. (2026), Table 1. CCP = correlation coefficient (r) and coefficient of efficiency (CE) against instrumental observations for GMT. CCSM4-based products shown for controlled comparison.

When compared against earlier paleoclimate data-assimilation products sharing the same parent climate model, LMR4D-Var comes out ahead. The earlier reconstructions — LMR Seasonal, LMR Online, and LMR v2.1, all CCSM4-based — achieve correlations of 0.90–0.93 and CE values of 0.77–0.80. LMR4D-Var with the CCSM4 emulator reaches r = 0.94, CE = 0.88, and the five-model mean tops out at r = 0.95, CE = 0.89 (Table 1 in the paper). Perhaps more important than the modest skill bump: LMR4D-Var adds long-memory observations and coupled ocean variables without sacrificing skill against annually resolved records.

The borehole influence shows up most clearly in the ocean. In the 300–2000-m layer, agreement with independent OHC estimates from Wu and colleagues jumps from CE = 0.19–0.20 without boreholes to 0.90–0.91 with them; detrended CE rises from 0.51–0.52 to 0.79–0.81. The same qualitative pattern holds against two other independent OHC reconstructions: against Zanna and colleagues, CE jumps from 0.12–0.13 to 0.80–0.81; against Gebbie and colleagues, from 0.18 to 0.60–0.61 (Meng et al., 2026). The borehole effect is much smaller in the 0–300-m layer, where the different configurations produce nearly identical trajectories.

Boreholes transform deep-ocean heat content skill

Label	Value
No boreholes (P2k)	0.2
With boreholes (P2k_BH)	0.91

Source: Meng et al. (2026), comparison with Wu et al. (67). BH = borehole assimilation. 'Detrended CE' removes the common long-term warming trend.

Perhaps the most elegant result comes from an out-of-sample check. The team assimilated only terrestrial boreholes, deliberately excluding ice-core borehole temperatures because accounting for ice advection would add considerable complexity. That exclusion made the West Antarctic ice-core borehole record an independent test. When the reconstructed temperature histories were passed through the borehole forward model at WAIS Divide, the root-mean-square error dropped from 0.052 ± 0.002 K in the PAGES2k-only run to 0.045 ± 0.003 K once boreholes were added (Meng et al., 2026). A single site doesn't constitute a formal significance test, but the pattern is telling: assimilating boreholes made the Little Ice Age reconstruction more consistent with a long-memory constraint that was never used in the optimization. The large-scale cooling is not merely an overfit to local borehole information.

The borehole assimilation also produces a La Niña-like LIA surface-temperature pattern with its largest amplitude over West Antarctica, and a cooler global-mean LIA ocean in both layers. This helps reconcile a long-standing tension in the literature: borehole-based reconstructions tend to imply greater Little Ice Age cooling than reconstructions built primarily from annually resolved proxies (Meng et al., 2026).

Why This Changes Things

The deepest claim here is methodological, and it ripples well beyond any single reconstruction. For two decades, the field has been operating with tools that structurally cannot combine the different temporal memories of different archives. Filters discard a borehole's centuries of memory when they treat it as an instantaneous update. Offline approaches lack an evolving dynamical memory altogether. LMR4D-Var is the first framework to treat every archive as the temporally extended witness it actually is, within a single coupled optimization, with explicit balancing of errors in model, observations, and initial conditions.

The ocean heat content result is the most consequential demonstration. OHC is arguably the hardest quantity in paleoclimate to reconstruct: it integrates energy uptake over decades to centuries while ocean circulation redistributes heat, and most proxies respond to surface conditions rather than subsurface temperatures. Yet the framework produces a dynamically constrained estimate of 0–300-m and 300–2000-m heat content that agrees strongly with independent estimates in the historical era — and the agreement is driven specifically by the long-memory borehole constraints, transmitted through the coupled emulator covariance. The authors frame the result directly: the borehole-related OHC improvement arises "naturally, without any explicit linkage between borehole proxies and OHC" (Meng et al., 2026).

The paper is also honest about where the framework falls short. The reconstructed multicentennial OHC amplitude is smaller than what Gebbie and colleagues estimate — which differs by roughly an order of magnitude in both layers. The likely culprit, the authors say, is the prior itself. The last-millennium model simulations used to train the emulators show small global-mean OHC variability, on the order of $10^8$ J m⁻², and contain strong long-term drifts, particularly in the deep layer. The emulators are trained after linear detrending, which removes unrealistic drift but may also strip out real multicentennial variability. The linear propagator estimated in truncated EOF space may not fully capture state-dependent changes in ocean circulation and heat uptake when strong forcing pushes the climate outside the training range. These approximations could attenuate OHC amplitude or distort its spatial distribution — but, the paper stresses, their relative contributions are not quantified here, and the historical-period trends remain close to independent estimates.

A hypothesis-generating contrast emerges from the layer-resolved maps. The Medieval Warm Period-to-Little Ice Age transition in deep OHC has a pronounced North Pacific component, whereas the modern OHC increase does not show the same spatial structure. The authors suggest the two intervals reflect different forcing regimes — volcanic forcing during the MWP-to-LIA transition versus anthropogenic greenhouse-gas forcing today — which can project onto different modes of ocean circulation and heat redistribution. They explicitly label this a hypothesis-generating result, not formal attribution, because targeted forced-simulation experiments are needed to test it.

What's Next

The Temp12k-only experiments point to the frontier. Temp12k records span decadal-to-millennial timescales, making them a natural test case for multiscale assimilation beyond the Common Era. In these runs, a model-error weight parameter $J_m$ controls the balance between the imperfect emulator and the low-frequency proxy constraints. Because the linear inverse model prior is trained mainly on last-millennium simulations, it does not fully capture Holocene-scale dynamics, particularly early-Holocene deglaciation. Reducing $J_m$ lowers the weight on the emulator and lets the trajectory respond more to the Temp12k observations.

As $J_m$ decreases, reconstructed global-mean temperature variability grows, particularly in the early Holocene, while staying broadly consistent with existing Holocene reconstructions — capturing the 8.2 ka event and the overall amplitude of Holocene temperature variability, a mid-Holocene thermal maximum near 6 ka, and a late-Holocene cooling trend. That late-Holocene cooling is notable because it differs from the reconstructions of Osman and colleagues and Erb and colleagues, which do not show the same global-mean cooling (Meng et al., 2026).

The authors are careful to counsel caution here. In the Temp12k-only setting, the emulator is known to have missing physics, and the balance between observational constraint and prior becomes delicate. This is precisely the kind of regime where a new framework's honesty about model error matters most — and where the weak-constraint formulation earns its keep, since it provides a principled dial for how much to trust the model versus the data across very long windows.

Where does this leave us? The immediate significance is that paleoclimate reconstruction now has a unified methodology that respects the physics of its own data sources. That matters because the stakes are not merely academic. Paleoclimate reconstructions define the envelope of natural climate variability, and that envelope is the yardstick against which modern and future change is measured. If the Little Ice Age ocean was cooler than previously estimated — if borehole memory reveals more ocean heat storage change than annually resolved proxies alone can see — then the baseline of natural variability shifts, and with it our calibration of what counts as extraordinary in the present.

The framework's design makes it extensible. Because it is differentiable and avoids per-archive adjoint models, it can absorb nonlinear, time-averaged, or history-dependent proxy system models as they are developed — a practical path toward incorporating the full physical process by which an archive records climate. And the Temp12k results hint at deeper-time applications across the Holocene and beyond, with suitable emulators.

The authors themselves keep expectations disciplined. "Even with these limitations," they write, "the historical-period OHC trends remain close to the estimates of Ref. (67), indicating that the framework retains skill for recent forced changes while leaving substantial uncertainty in multicentennial OHC variability" (Meng et al., 2026). The last sentence is worth sitting with: for recent changes, skill; for the deep past, honest uncertainty. That is the right posture for a field that is, at last, learning to treat every archive as the temporally extended witness it truly is.