Nakshatrala's description of the design trade-off is the most informative line in the account: "If you try to enforce one aspect, some other aspect will be violated," and the stated goal is to manage the errors of each chemical component down to something very small [9]. That is an honest account of what a physics-informed network is doing. The governing laws are not hard-wired into the solution; they are imposed while the model learns, and the model corrects itself when a prediction breaks them [4]. Satisfaction is approximate, which is why keeping outputs inside physically realistic bounds gets separate attention in the framework [10].
For a bulk species, an absolute error of a given size is a rounding matter. For a mineral present in trace amounts it crosses into impossibility, because a concentration below zero does not exist, and small computational errors can produce exactly that [7]. Mudunuru, an Earth scientist at Pacific Northwest National Laboratory on the project, puts the search itself as looking for a needle in a haystack [8]. So the positivity problem and the exploration target coincide.
The forecasting goal deserves a second read. The framework is meant to estimate things experiments do not supply: flow conditions, initial chemical distribution or abundance, reaction rates, permeability, porosity [11]. That is an inverse problem, and its premise creates its own verification difficulty. If a quantity cannot be measured, the only things available to test a forecast against are the sparse data used to fit it and the residual of the physics it was trained on [4] [11]. Neither excludes a different parameter set that matches the same data equally well.
The thing this account does not tell you is what the speed advantage costs. The comparison it offers is qualitative on both sides: conventional simulators model these processes accurately but demand significant computing resources and struggle when information is sparse [5], while data-driven models predict faster but need large training sets and can return physically unrealistic values [6]. No runtime, no error magnitude, no training-set size and no description of the test problems appears in the phys.org write-up [13] [14]. Forecast time and training time are separate budgets, and if a network has to be retrained for each new formation, training is the bill a user screening prospects would actually pay.
The narrow claim here is sound: constraining a learner with the flow and reaction laws is the right way to stop trace-concentration predictions from leaving the physically admissible range [4] [10], and unphysical output is the specific failure that has limited purely data-driven approaches [6]. The wider claim, that this helps locate critical minerals concentrated by fluids moving underground [12], runs entirely through parameters nobody can measure directly [11]. That link is what the numbers in Transport in Porous Media will have to carry [2].
