Physics-Informed Stochastic Neural Operators for Climate and Earth-System Prediction
Abstract
Abstract: Climate and Earth-system prediction requires simultaneous representation of nonlinear multiscale dynamics, uncertain initial and boundary conditions, incomplete observations, and physical conservation laws. Data-driven forecasting can capture large-scale statistical structure, yet long forecast horizons and rare events expose difficulties associated with physical consistency and calibrated predictive distributions. This paper develops a Physics-Informed Stochastic Neural Operator for Earth-System Prediction (PISNO-Earth), a proposed framework that maps observed spatiotemporal states and forcing fields to distributions of future climate states while explicitly penalizing governing-equation residuals and conservation violations. The methodology integrates a conditional neural operator, stochastic latent forcing, discretization-aware differential operators, and multi-objective estimation of predictive, dynamical and probabilistic losses. It specifies an observation-alignment protocol, temporal and geographic holdouts, ensemble evaluation, and stratified diagnostics for tropical cyclones, heavy precipitation and subseasonal anomalies. A closed-form diffusion problem supplies an independently reproducible numerical benchmark: at a normalized forecast horizon of 0.10, its leading-mode amplitude is approximately 0.3727 relative to the initial amplitude, illustrating how a conservation-compatible decay constraint can be checked without asserting an unmeasured forecasting gain. The proposed architecture makes physical consistency, uncertainty reliability, and resolution transfer jointly testable, rather than treating them as interchangeable performance criteria. The principal contribution is a falsifiable and implementation-oriented modelling specification with explicit ablations, evaluation metrics, and operational safeguards. Large-scale forecasting benefits remain hypotheses until the proposed system is trained and evaluated on independently sourced Earth observations and reanalysis archives. Keywords: neural operators; physics-informed learning; stochastic partial differential equations; climate prediction; Earth-system modelling; Bayesian uncertainty; probabilistic forecasts; conservation constraints; spectral operators; subseasonal prediction; extreme precipitation; ensemble calibration; resolution transfer; causal forcing.