Earth models that show their uncertainty: covariance maps, unseen wells, and climate surrogates

Inversion that maps its own limits

In a preprint, the useful output of an inversion is not only a subsurface image but a statement about where that image can be trusted. The stochastic variational inference framework for electrical resistivity tomography and time-domain induced polarization provides a computationally efficient alternative to Markov chain Monte Carlo by using adjoint-state gradients for both mean and covariance updates. Efficiency here is a practical claim: the same gradient machinery updates the central estimate and its spread. m:…3504

The recovered covariance in the framework serves as a spatial map indicating where geophysical data constrain the subsurface model. Where the map shows strong constraint, the data are doing the work; where it does not, the model is filling in. Field application at the Hongqi Reservoir demonstrates the ability of the framework to identify conductivity and chargeability structures associated with seepage outlets, so the map is tied to features with a physical reading, not only to abstract uncertainty. m:…3504

Groundwater at wells never trained on

A second preprint addresses a different kind of uncertainty: what happens at a location with no training record of its own. Foundation models adapted with low-rank adapters can accurately forecast groundwater levels at unseen wells without per-well training, according to the preprint. The test is direct, since the wells in question were not part of per-well training, and the model must transfer what it learned elsewhere. m:…5182

GeoAI-GWFM-KR, a time-series foundation model adapted to the Korean monitoring network, outperforms per-well long short-term memory networks in groundwater level forecasting at 7, 30, and 90-day horizons. The comparison spans short, medium, and longer horizons, so the result is not confined to one forecast window. Physics-informed loss functions did not improve the accuracy of per-well long short-term memory networks in this study, which is a useful negative result: adding physics-flavored training terms did not help the baseline it was tested on. m:…5182

A climate model tuned two ways

A third preprint makes model uncertainty explicit by calibrating one climate model to two different answers. E3SMv3 can be calibrated to achieve an equilibrium climate sensitivity of 2.69 K or 4.67 K using a machine learning surrogate model, while the default equilibrium climate sensitivity of E3SMv3 is 3.89 K. The spread between the two calibrated values is wider than the distance from either value to the default, which shows how much room calibration choices leave. m:…3052

A 400-member perturbed parameter ensemble, varying 14 atmospheric parameters, is sufficient to train a surrogate model consistent with 46 climate variables from satellite and reanalysis data. The ensemble size, the parameter count, and the variable count together describe the training basis for the surrogate. This suggests that the surrogate is not a shortcut around physics but a compressed map of many model runs, and that its value depends on whether that map stays consistent across the variables it was trained to match. m:…3052

Tools that move without retraining

A fourth preprint concerns tooling rather than a single model result, but it fits the same theme of stated fidelity. A four-step workflow enables the conversion of TensorFlow-based seismic phase pickers to PyTorch without retraining, preserving original model weights. Conversion without retraining matters because retraining can change behavior in ways that are hard to trace back to the original model. m:…4601

Converted EQCCT models, named EQCCTP and EQCCTS, exhibit mean absolute errors between 1.46 times ten to the power of negative nine and 1.82 times ten to the power of negative nine compared with original TensorFlow implementations. Those errors are small in absolute terms, and reporting them gives users a number for how much conversion changed the output. Across all four preprints, the pattern is consistent: the result is reported together with a measure of where it holds, where it weakens, or how far it moved. m:…4601