When an AI Solves the Physics: How Firm Is Each Answer? A Neural Operator Read at Several of Its Own Resolutions

Neural operators are AI models that learn the laws of a physical system, such as the flow of air, water or plasma. They answer questions about it far faster than classical solvers, and they are used for weather, fusion, chip design and medical devices. Each answer comes with no sign of how firm it is. This concept note reviews the main published work on neural operators and marks where an answer can go wrong without any sign. It covers the Fourier Neural Operator, the neural operator library, the Nature Machine Intelligence paper on extending architectures to function spaces, conformal prediction for operator learning, the discretization-error analysis of Fourier neural operators, and the ICLR 2026 study of zero-shot super-resolution. The note then describes a simple check for each case. A neural operator is built to run at any grid, so the grid is one of its own permitted settings. The computer makes 2 to 4 versions of the same run by moving only the model's own permitted settings: the normal setting and settings near both ends of the permitted range. The same trained model reads every version, unchanged. The spread between the versions is inversely proportional to the leading eigenvalue, so a wide spread means the answer is near a tipping point, such as a plasma disruption or the change from smooth to turbulent flow. A computer-decided shunting action then sends only the doubtful cases to a classical solver, and all other cases carry on. It needs no retraining and no new data. Patent pending. Keywords: neural operators, Fourier neural operator, scientific machine learning, resolution, discretization, tipping point, spread value, shunting action