The framework works in three stages. First, a Trust-Region Reflective (TRR) algorithm fits a bound-constrained exponential baseline to the historical record, establishing the broad downward trend. Second, a long short-term memory (LSTM) neural network, a type of recurrent deep-learning architecture suited to sequential data, learns the residual differences between the baseline and the actual observations.

Third, a physically constrained exponential refit incorporates the network's residual corrections back into a final exponential model, keeping the result consistent with known decay behavior. The reconstructed environmentally effective half-lives were 10.52 years for the Bohai Sea, 14.84 years for the Yellow Sea, 15.05 years for the East China Sea, and 16.40 years for the South China Sea. To test the model honestly, the researchers used a temporal holdout strategy: the network trained on data from 1960 to 2010 and was then evaluated on unseen measurements from 2011 to 2017.

Across the four seas, the residual correction reduced absolute prediction errors in all basins, with mean absolute errors between 0.30 and 0.61 becquerels per cubic meter of seawater. That very decline created a challenge for forecasting. Recent records are few in number and low in value, which makes single-number long-term projections unreliable.

The team therefore reports long-term projections as 90% ensemble prediction intervals rather than point estimates, an approach that preserves the uncertainty inherent in extrapolating from sparse, near-background data. Using a detection limit of 0.2 becquerels per cubic meter as the benchmark, the model projects that the Bohai Sea will fall below the threshold around 2036, followed by the East China Sea around 2049, the Yellow Sea around 2053, and the South China Sea around 2055. In other words, within roughly three decades, cesium-137 in China's surface coastal waters is expected to drop beneath routine detection.

By combining a physically motivated exponential model with a data-driven residual learner, the TRR–LSTM–TRR framework keeps its predictions anchored to decay physics while allowing the neural network to capture patterns that a simple curve fit would miss. The authors note that the claims are their own and do not necessarily represent their affiliated institutions, and the projections remain model-based estimates whose confidence depends on the quality and density of future measurements in these four seas.

Subject of Research: AI-based prediction of cesium-137 activity concentrations in the surface seawater of China's coastal seas

Article Title: A prediction model for 137Cs in surface seawater of China’s coastal waters based on AI oceanography method

Article References: Yu, Z., Men, W., & Wang, F. (2026). A prediction model for 137Cs in surface seawater of China’s coastal waters based on AI oceanography method. Frontiers in Marine Science, 13, Article 1924778. https://doi.org/10.3389/fmars.2026.1924778

Image Credits: AI Generated

DOI: 10.3389/fmars.2026.1924778

Keywords: cesium-137, AI oceanography, LSTM, Trust-Region Reflective algorithm, China seas, marine radioactivity, Bohai Sea, Yellow Sea, East China Sea, South China Sea, environmental half-life, seawater monitoring