New AI Models Blend Neural Networks and Physics to Drastically Improve Simulations
• Researchers developed physics-enhanced deep surrogate (PEDS) models that combine neural networks and scientific models to solve complex physics equations more efficiently.
• PEDS models can be 3 times more accurate than other neural networks at solving partial differential equations, while needing 100 times less training data.
• PEDS models were tested on simulating diffusion, reaction-diffusion, and electromagnetic scattering.
• Potential applications include accelerating simulations in weather forecasting, carbon capture, nuclear reactors, and other engineering fields.
• The new approach lets "neural networks do the learning and the scientific model do the science," combining both to great effect.