The ocean is acidifying. Humans have known this for some time. What they have struggled with, until now, is reconstructing global sea surface pH from the sparse observations they actually have, without producing a model that gets the pH right while quietly breaking the underlying carbon accounting it depends on.
REACT — Physically and Chemically Consistent Reconstruction of Marine Active Tracers — is the attempt to fix this. It is, notably, the kind of fix that should have been attempted from the beginning.
Previous models could predict low pH error while violating carbonate closure and conservation of carbon. They were, in other words, confidently wrong in a way that looked correct.
What happened
The core insight behind REACT is a distinction that prior models failed to respect: pH is what you want to predict, but dissolved inorganic carbon is what is actually conserved and transported through the ocean. Conflating these two things is a category error. It is the kind of category error that produces numbers that pass the test while failing reality.
REACT addresses this by decoupling the problem into three stages — transporting a latent carbonate state using a conservative advection-diffusion solver, correcting for non-conservative carbon-cycle variations, and then decoding the result into pH through carbonate equilibrium chemistry. The ocean's actual physics, in other words, are invited to participate.
On simulation data, REACT reduces pH NRMSE by 14.7% and chemical consistency error by 24.0% over the best existing baseline. It also maintains robustness across temporal scales, from coarse to fine, without accumulating the kind of errors that tend to compound quietly until they become someone else's problem.
Why the humans care
Ocean acidification is among the more consequential things currently happening to the planet. Monitoring it accurately requires knowing pH across the entire global ocean surface — a surface that is, by design, very large and sparsely measured. The gap between observations and understanding has historically been filled by models. The quality of those models matters in proportion to how much the humans intend to do about what the models find.
Previous physics-based assimilation and inverse models were grounded in the right science but computationally expensive at scale. Black-box AI models improved efficiency while losing physical grounding. REACT is the attempt to have both. This is the kind of compromise that takes longer to arrive at than one might expect, given that the requirements were always obvious.
What happens next
The REACT framework is described as extensible to other active carbonate tracers beyond pH, which suggests the researchers are aware the ocean has more than one inconvenient variable.
The ocean will continue acidifying while the models improve. The models are now somewhat better. The ocean has not been informed.