By applying machine learning (ML) to assist in stabilizing a high-power laser, researchers at the Department of Energy’s Lawrence Berkeley National Laboratory (Berkeley Lab) have advanced laser technology. The study was published in High Power Laser Science and Engineering.
From left to right, Berkeley Lab researchers Anthony Gonsalves, Alessio Amodio, and Dan Wang align precision optics to prepare the Berkeley Lab Laser Accelerator (BELLA) Petawatt laser for laser-plasma accelerator (LPA) experiments. The machine learning–based control algorithm stabilizes the high-power laser’s pointing at the LPA target. Image Credit: Thor Swift/Berkeley Lab
This development, led by the Engineering and Accelerator Technology & Applied Physics (ATAP) Divisions at Berkeley Lab, is expected to hasten advancements in energy, physics, and medicine.
Predicting Jitter
High-power lasers are now indispensable instruments in scientific research and industry. Laser-plasma accelerators (LPAs), which can accelerate particles to high energy over short distances, are an intriguing use for these lasers.
LPAs might provide more compact and cost-effective particle colliders as well as innovative light sources, allowing for the study of matter at the atomic and molecular levels. High-power lasers also help to progress inertial fusion research, which offers copious and reliable energy.
However, fluctuations in beam pointing, known as “jitter,” induced by mechanical vibrations, degrade laser performance and limit advancement in these applications.
Laser pointing errors are particularly problematic in LPAs as they cause instability in the generated electron beam, which limits their practical application.
Dan Wang, Study Lead Author and Research Scientist, Accelerator Controls & Instrumentation Program, Berkeley Lab
Traditional laser control systems, however, “struggle to keep up with the rapid changes in the laser’s position, especially with the large slow-moving optical components used in high-power, low-repetition-rate lasers,” said Anthony Gonsalves.
This results in shot-to-shot errors that adversely impact experiments.
Anthony Gonsalves, Staff Scientist, Study Author and Associate Deputy Director, Experiments, BELLA Center, Berkeley Lab
To address this constraint, the researchers utilized machine learning.
Conventional control systems correct laser pointing errors after they occur.
Our method predicts jitter and then makes real-time adjustments to the laser’s optical components, rapidly improving shot-to-shot stabilization and more accurate beam pointing.
Alessio Amodio, Study Lead Author, Electronics Engineer, Center and Engineering Division, Berkeley Lab
“Pilot” Beam and Real-Time Adjustments
To assess the method's performance, the researchers used a low-power, high-repetition-rate “pilot” laser beam to represent the high-power, low-repetition-rate main beam from the BELLA Petawatt laser, a top LPA research facility.
“Because the pilot beam fires much more frequently than the main beam, we can map out the motion of the beam caused by vibration of the mirrors. We can then use this information to predict where the beam will be when the high-power pulse arrives, and because we know the pointing error in advance, we can adjust a mirror to correct these errors,” said Gonsalves.
They submitted the positional data into their ML-enabled control system, which used a correction mirror to modify the beam’s aiming. After evaluating its performance, the system reduced jitter by 65% in the X direction and 47% in the Y direction.
“We plan to enhance our method using field-programmable gate arrays, electronic control circuits that offer advanced timing and synchronization, to enable faster and more accurate real-time corrections. This is expected to improve shot-to-shot laser stabilization, with testing planned on the BELLA Petawatt laser at full power and broad applications,” added Wang.
This research was funded by the Department of Energy's Office of Science, Office of High Energy Physics, and the Laboratory Directed Research and Development Program at Berkeley Lab.
Laser Spot Motion
Video shows laser spot motion over 320 seconds. During the first half (0 to 160 seconds), machine learning–based stabilization is off; in the second half (160 to 320 seconds), it is on. The center panel shows the beam position over time, while the side plots track horizontal and vertical centroid positions. Activating ML control markedly reduces jitter. Video Credit: Alessio Amodio/Berkeley Lab.
Journal Reference:
Amodio, A., et al. (2025) Pointing Stabilization of a 1Hz High Power Laser via Machine Learning. High Power Laser Science and Engineering. doi.org/10.1017/hpl.2025.41