Bayesian inversion trades the single best subsurface model for a set of weighted scenarios

Non-uniqueness is the whole reason this review has something to argue about. Because geophysical data are noisy, incomplete and unevenly informative, different underground models can predict very similar measurements ^[7], so a single best-fit image can look more confident than the evidence supports while the alternatives it edged out never reach the report ^[8]. The reframing the authors offer is the part with operational teeth: the test is not whether the picture is sharp, but whether the ambiguity left over could change a drilling location, a monitoring plan, or a safety assessment ^[9].

What a probability-weighted set of scenarios buys, on the authors' account, is a set of readable diagnostics rather than a verdict. Poorly constrained regions become visible, trade-offs between properties are explicit, different data types can be combined, and the risk of a costly misinterpretation can be estimated ^[13]. The same output separates the features the data strongly support from the ones that remain open, and those scenarios can be carried forward into predictions rather than collapsed first ^[11].

The enabling pieces come with conditions, and those conditions are what a review is built to weigh. Gradient-informed methods navigate large model spaces more efficiently by using derivatives, and that route is available when the simulator is differentiable ^[14]. Deep learning enters in two roles worth keeping separate: representing realistic geological patterns, and standing in for expensive physical simulations to support rapid repeated inference ^[15]. The first speeds up how quickly you can ask the question; the second widens which answers were ever available to find.

Differentiable Bayesian Inversion, as the authors describe it, is a framework that links geological knowledge, physical or learned simulators, measurement-uncertainty models and inference into one computational workflow, rather than a new sampling algorithm ^[16]. Plumbing usually decides whether a method leaves the literature, and this is plumbing. It is not a cost claim, and the interview does not make one: no compute comparison against a single best-fit inversion, and no named field case in which a posterior changed a decision ^[17].

So the shift the authors report, from finding one preferred model to evaluating a range of plausible models and predictions ^[12], is well argued as a change in what the deliverable should be and unmeasured as a change in what that deliverable costs. Where a decision turns on an unlikely and expensive scenario, and safe carbon dioxide storage is the clearest of the cases they name ^[5], the weighted set is the defensible product and a sharp image is an incomplete argument. Where the work is reconnaissance across a large, deep region that nobody wants to drill first ^[6], a single constrained description of the subsurface, physically consistent rather than a photograph ^[4], is still enough to tell you where to point the next line.