IBM's lunar model cuts ice-prospectivity RMSE to 0.0293 on a split without a geographic buffer

Patch-level assignment stops the same square of map from appearing on both sides of the split ^[18]. It does nothing about the square next door. The target makes that gap matter: Nyirjesy said the team generated ice prospectivity using the published workflow of Coyan and colleagues, 2025, which combines temperature, illumination, terrain and ice stability ^[11]. Those variables change slowly across a surface. A model trained on one patch is then scored on its neighbor, whose label was computed from much the same physics.

Asked how the evaluation keeps nearby regions from making the test easier, Nyirjesy said: "Neighboring patches were not explicitly separated by a geographic buffer, so the split prevents direct patch overlap but does not fully eliminate similarity between nearby lunar regions" ^[8].

Seed noise does not explain the result. The Table 7 difference is 0.0377 minus 0.0293, or 0.0084 RMSE, a 22.3% relative reduction ^[19]. That is 21 times the baseline's standard deviation of 0.0004 ^[20] and about six and a half times the fine-tuned model's 0.0013 ^[21]. Reproducibility across five seeds is a property of the training loop ^[5]. Of the two, the fine-tuned lunar model is the noisier, with roughly three times the seed spread of SwinV2-B ^[22].

Both models ran on the same predefined training, validation and test patches ^[9]. The missing buffer is therefore not an advantage handed to one side; if adjacent terrain makes the task easier, it makes it easier for the ImageNet-pretrained baseline too. Lets Data Science wrote that the absence of a buffer does not by itself establish that test data entered training, or that the measured improvement would disappear ^[14]. The open question is whether the 0.0084 holds on terrain that has no training patch beside it.

For the 22% to mean anything about a prospecting decision, both models would have to be retrained with contiguous regions held out, and the gap would have to survive the change. Lets Data Science said that is the evaluation it wants to see, and that Nyirjesy's answers do not include those results ^[15].

A regional split would still be scoring agreement with a modeled index. Nyirjesy said: "Ice prospectivity should be interpreted as an estimate of where conditions are favorable for near-surface water ice, not as a direct detection of ice" ^[10]. Establishing whether ice is present, she said, requires independent observations or direct measurements ^[12]. Repeated-run stability does not speak to that at all ^[13]. RMSE here summarizes the distance between predicted scores and the benchmark's target scores, weighting larger errors more heavily, and lower is better ^[17]. So the 22% is a statement about matching Coyan's index, not about finding 22% more ice ^[7].