Research map / Probabilistic and scientific modeling / Molecules and proteins

REACT: Physically and Chemically Consistent Reconstruction of Marine Active Tracers

NeurIPS 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Machine learning for physical simulation · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes REACT, a carbon-first framework that transports a latent carbonate state with an advection-diffusion solver, corrects it with a learned source module, decodes pH, and enforces carbonate-equilibrium closure, improving sea-surface pH reconstruction on MOM6-COBALTv2 simulations and WOD/SOCAT observations.

Problem

Reconstructing global sea surface pH from sparse observations while keeping it consistent with carbonate chemistry and source-free DIC conservation; existing physics-guided models treat pH as a passive tracer.

Approach

Transport a latent carbon (DIC-like) state with a conservative PDE rather than pH itself, add a learned active source correction for non-conservative carbon-cycle effects, decode to pH, and penalize carbonate-equilibrium inconsistency and source-free DIC mass drift.

Key results

• On month-to-week at 1% sampling, REACT reduces pH NRMSE by 14.7% and Chem. error by 24.0% over the strongest baseline (NUTS).
• On real WOD/SOCAT observations, REACT obtains pH NRMSE 5.25e-3 vs 5.68e-3 for 4DVarNet (about 7.6% reduction).
• Ablation on month-to-week at 0.11%: removing the source module raises pH error from 4.08e-3 to 5.79e-4 scale comparison; mass constraint reduces local drift from 8.84 to 2.26 per mille.

Surprising findings

• Adding the chemical constraint slightly worsens pH error while improving carbonate closure, a trade-off the authors call small.
• Removing pCO2 as an auxiliary input raises pH error roughly 3.7-fold, showing pCO2 is the most critical input.

Limitations

• Evaluated only on pH; extension to other active tracers left to future work.
• Chemical constraint trades slight pointwise pH accuracy for carbonate closure.
• Theoretical guarantees rely on identifiability and realizability assumptions (A1-A3).

Experimental setup

Datasets: MOM6-COBALTv2 simulation World Ocean Database (WOD) Surface Ocean CO2 Atlas (SOCAT)

Benchmarks: MOM6-COBALTv2 month-to-week MOM6-COBALTv2 week-to-day WOD/SOCAT real sparse observations

Metrics: pH NRMSE Chem. carbonate closure error Drift_local Drift_global

Baselines: Kriging (exponential) Kriging (spherical) 4D-Var Marble FNO U-Net ViT NUTS AtmoRep Prithvi WxC 4DVarNet

Builds on: NUTS 4DVarNet ViViT FNO

Closest papers by meaning

• OxyGenerator: Reconstructing Global Ocean Deoxygenation Over a Century with Deep Learning (ICML 2024 · similarity 0.85)
• CarbonSense: A Multimodal Dataset and Baseline for Carbon Flux Modelling (ICLR 2025 · similarity 0.83)
• Coupled Data and Measurement Space Dynamics for Enhanced Diffusion Posterior Sampling (NeurIPS 2025 · similarity 0.81)
• Analyzing and Guiding Zero-Shot Posterior Sampling in Diffusion Models (NeurIPS 2026 · similarity 0.81)
• Generating Full-field Evolution of Physical Dynamics from Irregular Sparse Observations (NeurIPS 2025 · similarity 0.80)
• Christoffel-DPS: Optimal sensor placement in diffusion posterior sampling for arbitrary distributions (NeurIPS 2026 · similarity 0.80)