Their study, published in Applied Physics B, shows that this self-correction produces measurably more uniform parallel machining on glass than either a purely computer-optimized hologram or an earlier self-correction scheme that relied on measuring the beam directly.
The technique centres on the computer-generated hologram, or CGH, a pattern displayed on a spatial light modulator that reshapes a single laser beam into many beams at once. In holographic laser processing, one pulse can be split into dozens of focused spots that all work on the material simultaneously, a route to higher throughput that the authors note is a constant industrial demand because processing speed is directly tied to cost. A CGH calculated on a computer can be extremely good, typically achieving a diffraction efficiency above 99 percent and an intensity error below 1 percent in simulation. The trouble begins when the pattern meets real hardware.
According to the researchers, the reconstruction degrades in an actual optical system because of environmental drift in temperature, air pressure and humidity, instability and aging of the laser and devices, device-to-device variation, undefined imperfections in the phase-modulation properties of the spatial light modulator, spatial variation of the light source, aberrations, slight misalignments and deposits in the optics. Fine-tuning the beam shape by hand is difficult and impractical at manufacturing sites. In-system CGH optimization was developed to address exactly this problem: the hologram is recalculated iteratively based on what the optical system actually produces, rather than what the computer predicts.
Previous demonstrations of in-system optimization had a fundamental weakness, however. They measured the reconstruction with a camera through a low-numerical-aperture lens in a space separate from the sample, and at low energies, to avoid damaging the imager. The authors point out that when an ultrashort laser pulse is tightly focused at high energy, its propagation changes through nonlinear optical effects. Conventional optimization therefore never saw the beam as the material sees it, and CGH optimization under practical laser-processing fluence conditions had not been achieved.
The new method replaces the direct beam measurement with laser-induced luminescence. When the femtosecond pulses are focused inside soda-lime glass, the material emits light through multiphoton excitation, and the brightness of that glow serves as an indicator of the local light intensity. Because the luminescence is generated on the sample itself, at the same laser fluence used for actual processing, and is imaged through the same objective lens that focuses the processing beam, the feedback signal faithfully reflects the beam state where it matters. The authors describe this as a step toward CGH optimization that accounts for nonlinear optical effects, something earlier methods could not do.
The experimental machine used a laser-diode-pumped Yb:KGW femtosecond laser with a maximum power of 10 watts, a centre wavelength of 1030 nanometres, a repetition rate of 10 kilohertz and a minimum pulse duration of about 155 femtoseconds. The beam was expanded to 10 millimetres and illuminated a liquid-crystal-on-silicon spatial light modulator. Telescope optics and an objective lens with a numerical aperture of 0.4 focused the diffracted beams to a calculated diameter of about 1.25 micrometres. A 700 to 900 nanometre bandpass filter in front of the observation camera suppressed surface reflections of the laser beam while letting the luminescence through.
Before using the glow as feedback, the team had to establish how it behaved. They found that no luminescence appeared below a laser fluence of 4 joules per square centimetre, that between 4 and 7 joules per square centimetre the luminescence depended linearly on fluence, and that above 7 joules per square centimetre it saturated. The authors attribute the linear regime to multiphoton absorption involving three or more photons, and suggest that void formation at the processing position contributes to the saturation under the multipulse conditions used for optimization. They carried out the optimization experiments in the linear range and confirmed that the dependence on diffraction angle, spatial-spectral wavelength dispersion and alignment error was small across the processing plane.
The optimization itself extends the weighted Gerchberg-Saxton algorithm, adjusting the amplitude weights of the hologram iteratively from the ratio of the target intensity to the measured luminescence, with 100 iterations run and the best result stored. Starting from a hologram that generated 38 parallel beams, the reconstruction uniformity, defined as the ratio of minimum to maximum beam intensity, had degraded to 0.40 in the real system despite a simulated value above 0.99. With luminescence-based feedback it climbed to 0.991. A parallel experiment using ordinary light intensity as the feedback signal raised the uniformity from 0.62 to 0.994, with a total runtime of 44.4 seconds for 99 iterations, or 0.45 seconds per loop, limited largely by the response speed of the spatial light modulator.
The practical payoff appeared when the three approaches were compared on the glass surface with 10-pulse irradiation. Counting the holes that could be machined as a function of pulse energy, the luminescence-optimized hologram produced a steeper curve, indicating more uniform parallel processing. At an average pulse energy of 24.2 nanojoules, the diameter uniformity, defined as the ratio of the smallest to the largest hole, was 0.433 for the computer-optimized hologram, 0.621 for light-intensity optimization and 0.773 for luminescence-based optimization, with standard deviations of hole diameter of 0.376, 0.206 and 0.126 micrometres respectively. Earlier, the team had also shown that the processed diameter correlates strongly with the normalized luminescence intensity, with a correlation coefficient of 0.93, supporting the glow as a meaningful proxy for machining outcome.
The method has clear limits that the authors state plainly. The sample must actually be processed while the beam is evaluated, so in these demonstrations the holograms were optimized in a preprocessing step inside the glass, with a time lag before surface processing, because internal processing quality is difficult to monitor in real time and surface processing is highly sensitive to pulse energy and focusing position near the threshold. They note that processing while observing luminescence should be possible through a warm-up operation, and that future systems will need monitoring designs tailored to individual applications. Even so, the demonstration shows that the light a laser leaves behind in the material can teach the optics that put it there to do a better job.
Subject of Research: In-system optimization of computer-generated holograms for parallel femtosecond laser processing using laser-induced luminescence as a feedback signal.
Article Title: In-system optimization of a computer-generated hologram using laser-induced luminescence for parallel femtosecond laser processing
Article References: Hayasaki, Y., Takahashi, K., & Hasegawa, S. (2026). In-system optimization of a computer-generated hologram using laser-induced luminescence for parallel femtosecond laser processing. Applied Physics B, 132(10), Article 104. https://doi.org/10.1007/s00340-026-08718-w
Image Credits: AI Generated
DOI: 10.1007/s00340-026-08718-w
Keywords: computer-generated hologram, femtosecond laser processing, spatial light modulator, laser-induced luminescence, holographic beam shaping, in-system optimization, multiphoton absorption, soda-lime glass, parallel laser machining, Gerchberg-Saxton algorithm, uniformity, Applied Physics B
News Source: Denise Maddox. (October 29, 2026). Laser-induced glow lets holograms tune themselves on the sample. Scienmag.
