An AI surrogate model is a machine-learning model trained to imitate the output of an expensive computer simulation. The exact simulation, for example a fluid-dynamics solver, a climate model, or a finite-element stress analysis, may take hours or days on a supercomputer. A surrogate learns from many runs of that simulation and then predicts its results in seconds, trading a small loss of accuracy for an enormous gain in speed.
The idea predates deep learning, but neural networks have made surrogates far more capable, especially through neural operators such as the Fourier Neural Operator, which can emulate the solution of an entire family of differential equations. Once a surrogate exists, tasks that require running the simulation thousands of times, such as design optimization, uncertainty quantification, or real-time forecasting, suddenly become feasible.
Surrogates are central to several headline results in AI for science. Machine-learned weather models replace parts of traditional numerical weather prediction with a learned emulator; digital twins of physical systems often rely on surrogates to stay responsive; and engineering teams use them to explore design spaces that would be impossible to search with full simulations.
For a business reader, the surrogate model is where AI for science meets the bottom line: the value is not a cleverer answer but the same answer arrived at fast enough and cheaply enough to use inside a tight design or decision loop.