Joseph Martin, & Dr. Alona Kryshchenko
“Laser-driven acceleration of protons presents exciting opportunities for future advancements in ultra-intense laser science and particle research. Optimizing laser parameters provides more control over the configuration of the laser system, producing a proton beam with an energy spectrum with desired characteristics. Invertible Neural Networks (INNs) are a recent development in deep learning and offer a promising approach to optimize these parameters. INNs are a class of normalizing flows with a computationally-efficient Jacobian determinant. This enables bidirectional training, reconstruction of input data from latent representations, and model interpretability. INNs have the potential to solve the inverse problem of achieving desired ion energy distributions by determining optimal laser parameters, and generate high-fidelity synthetic data through inverse transformations.
As a proof of concept, we developed and trained an INN on MNIST data to learn a model capable of forward and backward predictions. Our INN can construct unique images of handwritten digits, and accurately classify new images without retraining.”