Modeling laser-driven ion acceleration with deep learning
B. Z. Djordjević · A. J. Kemp · J. Kim · R. A. Simpson · S. C. Wilks · T. Ma · D. A. Mariscal
Abstract and summary · read the original at the source
In one page
A Lawrence Livermore group asked a practical question about the fastest ion beams made in a laboratory: when a short, intense laser pulse strikes a thin foil and blows a beam of ions off the back, which knobs actually set the ion energy? The honest way to answer is a particle-in-cell simulation, which follows the plasma particle by particle and costs so much computer time that sweeping the whole parameter space is out of reach. So Djordjević and colleagues ran a little over a thousand one-dimensional simulations and trained a neural network on them, producing a surrogate — a fast stand-in that maps laser intensity, pulse length, foil thickness, target density and the pre-plasma gradient scale length onto peak ion energy and hot-electron temperature. The surrogate answers in a heartbeat, so the whole space can be walked. What it turned up is the pre-plasma: ion energy peaks sharply when that thin front layer has a scale length near two tenths of a micron, a knob nobody was tuning.
Why it matters hereChapter 9 is about what plasma does when energy is poured into it faster than it can spread out, and laser-driven ion acceleration is where a laboratory reaches that regime today; chapter 12 keeps the ledger of the fusion routes those beams serve. The method carries as much weight as the physics — turning an expensive simulation into a surrogate you can turn like a dial is how a design programme finds its own optimum, and chapter 1 is where the site keeps the rules for trusting a result like that.
What it claims
01Peak ion energy and hot-electron temperature both track laser intensity most strongly, and then, in falling order, pulse duration, pre-plasma gradient scale length, target density and foil thickness — with the gradient scale length carrying more weight on ion energy than pulse duration does.Section IV C; correlation matrix, Figure 7(a)
Published and peer-reviewed02Ion energy peaks as a function of the pre-plasma gradient scale length near two tenths of a micron, a swing of more than ten mega-electronvolts, which points to a distinctive coupling between the laser and the plasma its own leading edge has already made.Sections IV C and IV D; Figures 7(g) and 9
Published and peer-reviewed03At fixed laser energy — intensity multiplied by pulse duration held constant — shorter and more intense pulses accelerate ions better than longer, weaker ones throughout the sub-picosecond range sampled here, and the energy gain flattens off once the pulse runs longer than about 150 femtoseconds.Section IV C; constant-energy curves, Figure 8(b) and 8(c)
Published and peer-reviewed04An ensemble of more than a thousand one-dimensional EPOCH particle-in-cell runs is enough to train a multilayer network into a working surrogate: the best models reached a mean-squared error of five parts in a million or less on outputs normalised to the range zero to one, in one to two hours of training on a single V100 graphics processor.Sections II A, III B and III C; Section IV A
Published and peer-reviewed05A surrogate can invent structure that is not physics: a streak the network drew in the intensity–duration map between 120 and 180 femtoseconds survived several retrainings and standard regularisation, then vanished once a denser block of 100 simulations was added — so features found this way have to be validated before they are believed.Section IV C; Figure 8(a) compared with Figure 8(b)
Published and peer-reviewed06The gradient-length feature is the result the authors want tested next, in higher-dimensional simulations, with shaped pulses, and out at higher intensities and thinner targets where other acceleration mechanisms — radiation-induced transparency, breakout afterburner, radiation-pressure acceleration — take over from sheath acceleration.Section V, conclusions
What to watch
Read it · abstract
Abstract
Developments in machine learning promise to ameliorate some of the challenges of modeling complex physical systems through neural-network-based surrogate models. High-intensity, short-pulse lasers can be used to accelerate ions to mega-electronvolt energies, but to model such interactions requires computationally expensive techniques such as particle-in-cell simulations. Multilayer neural networks allow one to take a relatively sparse ensemble of simulations and generate a surrogate model that can be used to rapidly search the parameter space of interest. In this work, we created an ensemble of over 1,000 simulations modeling laser-driven ion acceleration and developed a surrogate to study the resulting parameter space. A neural-network-based approach allows for rapid feature discovery not possible for traditional parameter scans given the computational cost. A notable observation made during this study was the dependence of ion energy on the pre-plasma gradient length scale. While this methodology harbors great promise for ion acceleration, it has ready application to all topics in which large-scale parameter scans are restricted by significant computational cost or relatively large, but sparse, domains.
B. Z. Djordjević, A. J. Kemp, J. Kim, R. A. Simpson, S. C. Wilks, T. Ma and D. A. Mariscal, Modeling laser-driven ion acceleration with deep learning, Physics of Plasmas 28, 043105 (2021), from Lawrence Livermore National Laboratory, with the University of California San Diego and the Massachusetts Institute of Technology. The accepted manuscript is LLNL-JRNL-817657, dated 13 December 2020.
(Abstract only — see the rights note above. On this site, ion acceleration seen directly in a magnetic fusion experiment is at /library/stm-5bd599dfca, the state of laser-driven fusion is at /library/stm-f585d315d5, and what extreme laser intensities do to quantum systems is at /library/stm-3b9cee0ab4.)
The way in
https://doi.org/10.1063/5.0045449The published article is Physics of Plasmas 28, 043105 (2021). Crossref records only AIP’s rights-and-permissions page as its licence and no Creative Commons statement appears, so the AIP article is not open-licensed. The text read for this sheet is the authors’ own accepted manuscript, released by Lawrence Livermore National Laboratory as LLNL-JRNL-817657 through OSTI. That manuscript carries the standard Department of Energy contractor disclaimer for work performed under Contract DE-AC52-07NA27344 and no public-domain dedication or Creative Commons statement, so it does not promote the sheet to open licence. The abstract below is the published one; every claim is located to a section or figure of the LLNL manuscript.
How to cite it
B. Z. Djordjević, A. J. Kemp, J. Kim, R. A. Simpson, S. C. Wilks, T. Ma, D. A. Mariscal (2021) Modeling laser-driven ion acceleration with deep learning. doi:10.1063/5.0045449
Where it sits in the curriculum
Plasmoids, charge clusters and the orbsLattice confinement fusionThe evidence ladder