Training Neural Networks on Synthetic Data for Target Normal Sheath Acceleration

Neural networks and other machine learning algorithms are beginning to be used in ultra-intense laser physics. An important concern is determining the minimum amount of data points needed for a neural network to be a viable approximation function because many laser systems are highly limited in how many shots per day they can operate. By using synthetic data based off of a model proposed by Fuchs et al. 2006, this paper explores the performance of a neural network with one hidden layer. When trained on datasets of varying sizes and for various numbers of epochs our work suggests that a minimum of 20000 data points are needed and that at least 30 epochs are needed to make a viable neural network that can estimate the max and average proton energies at a relative error below 10% and the total proton energy at a relative error below 20%. The results of this study can potentially be applied to training neural networks on real experimental datasets. A future direction for this research will be to train networks on a non-uniformly sampled parameter space in a way that is similar to how a real laser system would collect the data.

Fuchs et al. Nature Physics volume 2, pages 48–54 (2006)