Using long short term memory (LSTM) neural networks allows for better control of motorized alignment components in high-power laser accelerators, by integrating hysteresis in the beam correction feedback system. In the white zones (hysteresis correction off), pointing errors are larger than in the blue zones (correction on).

Improving the stability and reliability of laser plasma accelerators is paramount to their adoption in future high impact applications. At BELLA, we use AI and ML to help solve this problem in a number of ways, including improving feedback and control loops and identifying “hidden” parametric dependencies. 

The laser systems at the BELLA Center typically start with a low power laser firing 1000 pulses per second. A few of those pulses each second get amplified to 1000 times higher energy. Only the high energy pulses are used for laser-plasma interaction experiments. This unique pulse structure allows new approaches to implementing feedback and stabilization systems; for example, training neural nets on the environmental conditions in the lab can yield predictive insight into the stability of the laser facility and how to address mitigation strategies.

Left: Two images of a deflected probe laser after propagation through a plasma (accelerating) structure, indicating at first glance very little variation. However, neural networks can reveal a strong correlation between these kinds of images and the accelerated electron beam charge produced. Right: Comparison between the measured beam charge and the neural-network predicted charge highlights the potential of the network’s capability.

Furthermore, correcting the pointing of the high power laser systems often requires constant monitoring and control. In some instances, hysteresis of motor actuators can limit the accuracy of corrections. Modeling the hysteresis using long short term memory (LSTM) neural networks has recently proven effective in improving overall stability, reducing pointing fluctuations by more than a factor of 2.

The process of generating and accelerating particle beams in plasma-based accelerators involves numerous complex nonlinear processes. Understanding the sensitivity of the interaction to small variations in laser and plasma parameters can be difficult. Interferograms, like those shown in the figure below, are often used in combination with complex post-processing algorithms to characterize a plasma profile. The complex and rich plasma profile images, however, can contain additional information not captured in the reduced post processed data. A simple convolutional neural net trained on sets of raw interferograms has been able to accurately predict the charge of LPA-produced electron beams.