At Singapore University of Technology and Design, researchers have trained an artificial intelligence to do what computer simulations have struggled with for years: design optical surfaces that work in the real world, not just in theory.
The breakthrough addresses a frustration that has long plagued optical engineers. Designing nanostructured surfaces called Fourier gratings—used in everything from compact spectrometers to augmented-reality displays—has relied on computer simulations that assume perfect conditions: smooth surfaces, single angles of light, zero imperfections. When these devices are actually manufactured and tested, they behave differently. The gap between simulation and reality has narrowed innovation.
Associate Professor Dong Zhaogang and his team, collaborating with researchers from Xiamen University and Hefei University of Technology, bypassed simulation entirely. Instead, they trained a transformer-based neural network called ExpForm directly on real experimental measurements. The team used a high-throughput spectroscopy system to collect over 25,000 spectral readings from fabricated nanostructures, capturing the messy reality: surface roughness, structural asymmetries, measurement noise. Each measurement took roughly six minutes. This raw, imperfect data became the AI's teacher.
The result is dramatic. ExpForm achieved 99.79% consistency with actual experimental measurements while evaluating optical spectra approximately 900 times faster than conventional simulations. Where traditional simulation failed to capture key spectral features under oblique and angled light, the AI-trained model succeeded. The difference is not academic—it collapses design cycles from hours or days down to seconds.
The framework works bidirectionally. A forward network takes structural dimensions and light angles as input, then predicts the resulting optical spectrum in real time. An inverse network reverses the process: given a desired spectral response, it identifies which structural dimensions and illumination angles are needed to produce it. Together, they replace the grueling traditional cycle of simulate, fabricate, measure, and repeat. For researchers and engineers, that means rapid prototyping without fabrication trial-and-error.
The inverse design capability opens possibilities previously considered too complex to solve. The team demonstrated on-demand generation of narrowband resonances at specific wavelengths, high-reflectance profiles, and dual-band resonances—all by adjusting the angle of incoming light rather than refabricating physical structures. One illumination angle can activate different optical modes without touching the geometry. "This effectively introduces an additional degree of freedom beyond geometry, expanding the design space significantly," Dong said. Yet practical use of this parameter has been limited by computational instability in simulations.
The paper, published in PhotoniX, reveals how artificial intelligence can bridge the chasm between idealized theory and physical reality. By training on real fabricated samples—complete with their imperfections and irregularities—the AI learned patterns that traditional physics-based simulations miss. The approach promises to accelerate not just optical design, but any nanoscale engineering where simulation and reality diverge. It is a reminder that sometimes the most powerful computational tool is not a better simulation, but learning directly from the world as it actually is.