Lawrence Berkeley National Laboratory, United States of America
Deep learning methods show great promise, however, applications where simulation data is expensive to obtain do not lend themselves easily to applications of deep learning without incurring a high cost to produce data. Real-time online learning is a novel strategy to minimize this cost: a model "learns as we go", only requesting additional (expensive) data if it encounters a situation where it needs additional training. We demonstrate the feasibility of this approach by accelerating a partial differential equation solver as it runs by training a convolution neural network in real-time to propose initial solutions which are optimized to accelerate the solver. To overcome typical challenges associated with online training of neural networks we introduce a methodology to selectively construct the training set as more data becomes available. We present results on physically motivated test problems that demonstrate the acceleration achieved using this real-time deep learning methodology.