Fig. 1: Experimental schematic and model process.
From: A neural network-based synthetic diagnostic of laser-accelerated proton energy spectra
The main laser pulse is transported via multiple stages of the laser beamline and finally turned with a dielectric mirror onto a F/2 parabolic mirror that focuses the pulse onto a tape target, producing a plasma and driving ion acceleration. Laser light back-reflected from the plasma, together with second harmonic light generated in the plasma, passes through the dielectric mirror and onto a scatter screen, where the spatial, intensity and spectral properties are measured. The spectrum of the beam of protons accelerated from the plasma is measured using a Thomson parabola ion spectrometer. As shown in the dashed inset, a lower intensity preheater laser pulse is focused onto the target at time Δt prior to the arrival of the main pulse. The target displacement with respect to the main laser pulse focus, Δz, is also varied. The model training and prediction processes, and example spectra, are shown on the right. The β-VAE is trained to encode and reconstruct (blue) the measured (black) proton spectra, after which the encoder is used to generate a mean latent space representation for all spectra in the training set. The predictor neutral network is trained with the laser parameters and back-reflection moments as inputs to predict the mean and standard deviation of the encoded proton spectra, by minimizing the negative log-likelihood loss function (LNLL). The mean latent space prediction is decoded by the β-VAE decoder to produce the predicted spectrum (green), with 1000 samples from the predicted latent space distribution decoded to generate a predicted distribution (each sample shown in a shade of red with reduced opacity to visualise the distribution of overlaid data).
