Lawrence Livermore National Laboratory

Sandrine Herriot


The need for a new concept for laser control is driven by the rapidly growing field of high-energy, high-repetition-rate, short-pulse laser development worldwide. Efficient use of such lasers requires leveraging an active feedback control loop to optimize the laser performance directly through the outcome of an experiment. Such optimization entails the ability to safely scan the laser's parameter space and rapidly evaluate the risk of damage (to the laser itself) associated with the configurations proposed by the optimizer in the loop. Conventional approaches using analytical models present limitations to achieve the required speed for those next-generation kilohertz-repetition-rate lasers; thus, the goal of this project was to develop a new concept to enable fast (kilohertz-rate) and accurate assessment of a laser's performance and the associated risk of damage to chirped-pulse-amplification (CPA) laser systems, applicable in a feedback control loop.

Our approach is based on a deep-learning neural network (NN) used as modeling code for CPA laser systems and associated optical damage physics. NNs are able to learn from labeled data and generate a surrogate model of the phenomena under study. We focused on developing a capability for quickly modeling CPA laser performance, in particular CPA amplifiers. For that, we first developed an approach using a simplified one-dimensional (1D) analytical code that demonstrated limited speed (tens of hertz), causing a bottleneck for our kilohertz-regime lasers; then we developed a NN algorithm as an alternative to the 1D model of the CPA amplifier and demonstrated significant speedup (tens of kilohertz) compared to the analytical model capability, along with excellent accuracy.

The laser-induced optical damage is a critical aspect to account for when running an automated optimizer loop, especially in a short-pulse regime at a high-repetition rate. We therefore developed a convolutional neural network algorithm (CNN) to model the damage physics considered in our system, which is based on the pulse shape (Tien et al. 1999) and demonstrated its ability to predict the risk of damage associated with different laser configurations in terms of speed and accuracy, with less than 1% false safe predictions.

The capacity of the model to adapt itself to change in the laser performance is another problem that needs to be addressed. We demonstrated the capacity of our NN-based code to calibrate itself over time by adding newer labeled data from the laser to the training set, thus providing reliable predictions. We observed that the number of labeled data and their diversity have a strong impact on the generalizability and accuracy of the NN-generated surrogate model. This led us to develop a Bayesian deep-learning code for an active-learning approach enabling efficient and fast training of the NN, which therefore required fewer labeled data, as opposed to a non-active-learning approach.

Ultimately, our approach, which entails using machine learning as a modeling tool for laser performance, is key to addressing the speed required by future kilohertz-rate laser systems that employ a feedback loop. This research has laid the groundwork to address the modeling and analysis challenges raised by the goal of efficient use of the next-generation high-repetition-rate short-pulse lasers.

Background and Research Objectives

The development of high-repetition-rate, high-energy, short-pulse lasers is an emerging field with growing interest and investment worldwide (Haefner et al. 2016). This trend is driven by the potential applications for which petawatt-class CPA lasers with increased repetition rate can be used (Haefner et al. 2016, Siders and Haefner 2016) in industry, medicine, and national security, as well as high-energy-density science and high-field physics. Indeed, high-repetition-rate, high-peak-power laser systems will enable unprecedented, high-fidelity data through improved statistics and will address applications that require higher secondary particle or photon flux (González et al. 2018, Rothhardt et al. 2016).

Taking full advantage of these new capabilities will involve altering laser configurations very rapidly, even allowing for experimental results to provide feedback to adaptive control loops that optimize a particular experimental outcome by adjusting laser parameters in real time. Such a real-time active feedback control loop raises multiple challenges on the diagnostic, analysis, and control fronts. Any adaptive control loops that define the laser configuration must be constrained to avoid laser-induced optical damage to the laser itself (Tien et al. 1999, Strickland and Mourou 1985), and there is insufficient time at kilohertz rate between laser pulses to model the proposed laser configuration in sufficient detail to predict the onset of damage with high fidelity. The development of new concepts to address those challenges is critical to unlock the full potential of the next generation of high-repetition-rate, short-pulse lasers.

We developed  a new concept for laser modeling to replace the high-fidelity analytical model traditionally used in a single-shot laser system such as the National Ignition Facility to assess the risk to the laser of a given configuration. The latter is unfortunately time- and resource-consuming and hardly applicable to high-repetition-rate lasers. Therefore, our goal was to develop a novel machine-learning algorithm with several objectives derived from the requirements set by an active feedback loop control for kilohertz repetition rate. Our algorithm meets all four objectives. It can

  1. accurately model the laser physics involved in a CPA amplifier such that it can reliably infer its performance for any configuration of the laser within its parameter space;
  2. model the physics of laser-induced damage involved in CPA amplifier systems in order to accurately assess the risk to the laser associated with any laser configuration that an optimizer may propose within a feedback loop scheme;
  3. perform its evaluation of the laser performance and its risk factor fast enough, allowing kilohertz repetition rate operation; and
  4. update itself in real time to a variances in the system, such as drift and fluctuation, so that its assessment stays accurate and reliable over time.

Impact on Mission

Development of next-generation high-average-power, high-intensity lasers are of growing importance worldwide and are a key capability to address NNSA missions. Our machine-learning–based approach to laser performance modeling is a key component in helping to unlock the full potential these lasers and providing users with a novel approach to performing experiments. To the best of our knowledge, this is the first demonstration of high-energy, short-pulse CPA amplifier modeling using machine learning. This new modeling concept enhances the Lawrence Livermore National Laboratory's core competencies in lasers and optical science and technology, as well as high-energy-density science.


This research was motivated by the scientific need to exploit the full potential of next-generation high-repetition-rate, high-peak-power laser systems in a safe manner during the automated optimization process of a laser-based experiment. We leveraged deep learning to develop a real-time laser modeling capability at kilohertz rate, with self-calibration. A Bayesian neural network was developed to filter unsafe laser input configurations to safeguard against damage. The model can predict the laser amplifier output power spectrum with high accuracy for a wide range of input power spectra, with pulse energies spanning several orders of magnitude and with a variety of spectral shapes. Similarly, damage risk can also be predicted rapidly and accurately. In addition, the model provides uncertainty estimates that are used to (1) detect laser failures that are indicated when the ground truth measurement falls outside acceptable error bounds and (2) efficiently augment the training data pool to reduce model uncertainty, which in turn generates data that can lead to better coverage of the model’s understanding of the input space. The model also accounts for any change or drift occurring in the laser, so its predictions remain reliable at any time. Our results show that a machine-learning-based model is much faster than the physical model and can predict with accuracy repetition rates exceeding, respectively, 98% and tens of kilohertz. Our research is a first step toward revolutionizing the way high-energy, high-repetition-rate, petawatt-class lasers are currently used to perform experiments. Our approach demonstrates the possibility of achieving real-time safety control and optimization of these next-generation, short-pulse laser systems, which is a necessary capability for maximizing their utility for any laser-based experiments.


González, A. I., et al. 2018. "High Photon Flux XUV Source Driven by High Repetition Rate > 100 kHz Fiber Laser." Proceedings of High-Brightness Sources and Light-Driven Interactions, OSA Technical Digest.

Haefner, C. L., et al. 2016. "Next Generation Petawatt Laser Systems." 5th Advanced Lasers and Photon Sources Conference (ALPS’16), Yokohama, Japan, 17–20 May, 2016.

Rothhardt, J., et al. 2016. "High-Repetition-Rate and High-Photon-Flux 70 eV High-Harmonic Source for Coincidence Ion Imaging of Gas-Phase Molecules." Optics Express 24: 18133–47.

Siders, C. W., and C. Haefner. 2016. "High-Power Lasers for Science and Society." White paper presented to the National Academy of Science Committee on the Opportunities in the Science, Applications, and Technology of Intense Ultrafast Lasers, 5 Oct., 2016. LLNL-TR-704407.

Strickland, D., and Mourou, G. 1985. "Compression of Amplified Chirped Optical Pulse," Optics Communications 56: 219–21.

Tien, A.C., et al. 1999. "Short-Pulse Laser Damage in Transparent Materials as a Function of Pulse Duration." Physical Review Letters 82: 3883.

Publications and Presentations

Herriot, S., et al. 2017. "Active Adaptive Control of High Energy, High Repetition-Rate, Short-Pulse Lasers." Lawrence Livermore National Laboratory, Livermore, CA, November 2017. LLNL-POST-747330.