A deep learning model can simulate certain interactions between laser pulses and crystals in milliseconds, according to a report on work by researchers at Stanford University, the University of California, Los Angeles, and SLAC National Accelerator Laboratory. The approach could speed up calculations used to understand and tune powerful laser systems, including those that help generate X-rays for scientific experiments.

The report, based on material provided by SPIE—the International Society for Optics and Photonics—describes the work as a way to model second-order nonlinear optical processes, also known as χ² processes. In these interactions, light waves exchange energy as they travel through specially designed crystals. The process can change light’s frequency and help control the shape of laser pulses.

That control matters at SLAC’s upgraded Linac Coherent Light Source, or LCLS-II. There, infrared laser pulses are converted first to green light and then to ultraviolet light. The ultraviolet pulse strikes a cathode and releases electrons, which are accelerated and manipulated to generate X-ray pulses. The timing and shape of the ultraviolet pulse help determine the characteristics of the electron bunch, which in turn influence the X-rays used in experiments.

Predicting these conversions requires substantial computing. Conventional simulations use the nonlinear Schrödinger equation to calculate how light changes as it moves through a material. A widely used approach, the split-step Fourier method, repeatedly switches between descriptions of a pulse in time and in frequency. The report says that this calculation accounts for about 95 percent of the processing time in the complete laser simulation examined by the researchers.

To address that bottleneck, the team adapted a long short-term memory network, a type of recurrent neural network, to handle interactions among multiple light waves. Instead of repeatedly converting between time and frequency representations, the model works with a compressed representation of the light’s frequency characteristics. The researchers tested it on noncollinear sum-frequency generation, a process in which three interconnected light fields evolve together inside a crystal.

The tests covered a wide variety of pulse shapes, including challenging cases with pronounced gaps in their frequency spectra and strong variations in optical phase. The model reproduced changes in the pulses over time and across frequencies. When the team processed multiple simulations together on graphics processing units, the average calculation time per simulation fell to a few milliseconds. The report describes the speed improvement as orders of magnitude over conventional methods, without substantially compromising accuracy.

The model also predicted the linked behavior of the interacting light fields: when it accurately predicted the main output of the sum-frequency process, its predictions for the other fields were generally accurate as well. That result suggests the model captured features of the interaction as a whole, though the report does not give an absolute error measure for its predictions.

The researchers see faster simulations as a possible step toward more responsive laser systems. Their modular design could represent separate physical processes with separate trained networks, which might eventually be combined to model an experimental system while it runs. They also describe possible uses in digital twins, adaptive laser control, and integration with diagnostic instruments.

Those applications remain prospective. The reported work tests a model against simulations; the account does not establish that it has been connected to an operating laser or used to control one in real time. If the approach can be extended and validated for such systems, faster predictions could help researchers monitor changes and adjust laser settings more quickly. The paper, “Deep learning-assisted modeling for χ(2) nonlinear optics,” is listed in Advanced Photonics, with Jack Hirschman, Erfan Abedi, Minyang Wang, Hao Zhang, Abhimanyu Borthakur, Justin Baker, Andrea L. Bertozzi, Randy Lemons, and Sergio Carbajo named as authors.