Researchers have used machine learning to optimize the output of a plasma accelerator in which intense laser pulses (sent through the metal cone on the right) ionize and set up very strong electric fields within a gas (contained in the metal cube). [Image: Rob Shalloo]
Devices that use intense laser pulses to set up very strong electric fields inside plasma hold the promise for a new generation of very compact and low-cost particle accelerators. But these “laser wakefield accelerators” are also finicky, needing to be very finely tuned to maximize their output.
Researchers in the U.K. have now shown how to overcome that problem by exploiting machine learning to optimize the input parameters of such an accelerator. The fully automated scheme, they say, should prove a boon for non-expert users who require energetic electrons or X-rays for their research (Nat. Commun, doi: 10.1038/s41467-020-20245-6).
The promise of laser wakefield accelerators
Laser wakefield accelerators fire very brief and intense pulses of laser light into a plasma to set up a wave of positively and negatively charged regions in the pulses’ wake. Electrons either from within the plasma itself or injected from outside ride this wave and in doing so can be accelerated to energies of several gigaelectronvolts in just a few centimeters.
The technology has the potential to replace much longer radio-frequency machines in accelerating electrons for a range of applications in science and beyond. The accelerated particles could, for example, be used to generate X-rays for medical imaging or even perhaps reach the energies needed to build a cut-price linear collider for particle physicists.
To reach peak performance, these devices rely on fine-tuning a large number of input parameters describing both the laser pulse and the plasma. However, many of these parameters are interdependent, with a slight change to one often significantly altering the value of others. That makes it very hard to optimize an accelerator’s output by adjusting each input in turn.
Adjusting parameters
In the latest work, Rob Shalloo of Imperial College London (now at the DESY lab in Germany) and colleagues have used what is known as Bayesian optimization to adjust as many as six parameters at the same time. They have shown that this allows a plasma accelerator to converge onto an optimum set of input values in a short space of time.
The optimization process involves creating a model that describes how the value of a particular output parameter, or “objective function,” changes as a certain set of input parameters is varied. The idea is to use a relatively limited set of experimental data to bring the modeled version of this function in line with that of the real world—and in doing so identify where in its multi-dimensional space the output is maximum.
Shalloo and co-workers put their technique to the test on a titanium-sapphire laser at the Rutherford Appleton Laboratory in Oxfordshire, U.K. They fired 245-mJ, 45-fs pulses at a small gas target to generate electrons and X-rays with energies of millions and thousands of electron volts, respectively.
Employing the algorithm
The researchers used their algorithm to boost the accelerator’s output in a series of completely autonomous steps. In each cycle, the program calculates new values for the input parameters, such as the laser’s spectral and spatial phase as well as the electron density and length of the plasma. It uses these values to adjust the relevant components (an acoustic-optic filter and deformable mirror in the case of the laser), before firing a series of laser pulses at the plasma and measuring the output via a magnetic spectrometer and X-ray camera.
Using the relevant objective function, the algorithm updates the modeled function and then selects new input values that are either close to a peak or in a relatively unexplored region of the parameter space. In this way, it can fairly quickly work out where the global maximum of the function lies.
The researchers tested their technique against several functions. For the simplest of these—electron spectrometer counts—they used four inputs and found they could reach the maximum output after just 20 iterations, which took 6.5 minutes. Optimizing the X-ray yield instead took nearly half an hour, requiring as it did some 50 iterations with six inputs. But the wait was worth it, generating a five-fold increase over the previous day’s yield and an actual X-ray image—of an array of materials placed in the path of the radiation—even though, they say, the laser’s energy would normally be considered inadequate for such imaging.
A tailored approach
In addition, they showed that by carefully selecting specific objective functions they could tailor the output of their automated accelerator to different tasks. In one test they targeted the electron beam’s total energy—key to generating positron beams or gamma rays—while in another they instead set out to limit the beam’s divergence, important when it comes to hooking up multiple accelerating stages. That targeting worked, given that electron bunches produced by the former beam packed more than three times the energy of those in the latter but were far more spread out in space.
The researchers also demonstrated just how important very subtle changes in the beam parameters can be in determining the accelerator’s output. In one test they were able to boost the beam’s charge by 80% even though the pulse length changed by only 1%.
Group member Matthew Streeter of Queen’s University Belfast in Northern Ireland says that the researchers next plan to test their algorithm by using it to automate experiments with electron and X-ray beams for a variety of applications. He envisages plasma accelerators becoming a standard tool across fundamental and applied science, explaining that their technique could potentially be used in any facility designed with centralized control and automation in mind. “Adding the extra software layer to provide the artificially intelligent control is a relatively straightforward thing to do,” he says.