Researchers at MIT and IBM have developed a new, more efficient method for creating data-driven surrogate models to solve complex physics problems. These models, which can be trained on a set of data from high-fidelity numerical solvers, are designed to predict the output of the partial differential equations (PDEs) for new inputs. This method, called “physics-enhanced deep surrogate” (PEDS), combines a low-fidelity physics simulator with a neural network generator. The researchers found that PEDS can be up to three times more accurate than an ensemble of feedforward neural networks with limited data, and it can reduce the training data needed by at least a factor of 100 to achieve a target error of 5 percent.
In addition, the PEDS framework has the potential to make minimal models more accurate and predictive for surrogate model applications. Complex physical systems governed by PDEs, such as climate modeling and seismic modeling, could greatly benefit from this new approach. The research was supported by the MIT-IBM Watson AI Lab and the U.S. Army Research Office.