Researchers from Singapore and China have developed a deep learning framework that trains artificial intelligence directly on experimental measurements to design optical Fourier surfaces, effectively closing the gap between theoretical simulations and real-world fabrication outcomes. The breakthrough, led by Associate Professor Dong Zhaogang of the Singapore University of Technology and Design alongside Professor Zhu Jinfeng of Xiamen University and Dr. Wei Chen of Hefei University of Technology, was recently published in the journal PhotoniX. Optical Fourier surfaces manipulate light for compact spectrometers, augmented reality displays, and advanced sensors. Traditional design workflows rely on computational simulations assuming idealized conditions, such as single-angle illumination and perfectly smooth geometries. These assumptions frequently diverge from experimental reality, particularly when leveraging incident light angle as a functional parameter. Simulations at oblique incidence are computationally expensive and unstable, while physical systems inherently involve angular distributions and fabrication imperfections. To overcome these limitations, the team engineered ExpForm, a transformer-based neural network trained exclusively on empirical data. The researchers deployed a high-throughput spectroscopy system to capture over twenty-five thousand broadband reflectance spectra from nanoimprint-fabricated samples. Each measurement incorporated real-world variables, including structural asymmetry, surface roughness, and instrumental noise. By bypassing first-principles simulation entirely, the model internalizes physical tolerances and actual optical behaviors. ExpForm operates through dual pathways. A forward network predicts spectral output in real time based on structural and angular inputs, while an inverse network computes the precise dimensional and illumination parameters required to achieve a target optical response. This bidirectional architecture replaces the conventional iterate-and-refabricate cycle with instantaneous computational design. Benchmarked against finite-difference time-domain simulations, the AI model demonstrated ninety-nine point seven nine percent consistency with experimental data and accelerated spectral evaluation speeds by approximately nine hundred times. The practical implications for nanophotonics engineering are substantial. Design cycles previously measured in hours are compressed to seconds, drastically reducing prototyping costs. The inverse capability also enables angle-programmable devices, where a single surface dynamically switches between spectral responses simply by altering illumination geometry. The team has publicly released the complete training dataset to facilitate cross-institutional benchmarking and accelerate industry adoption. This methodology marks a structural shift from theoretical modeling toward data-informed, experimentally grounded co-design. While prior experiment-driven machine learning approaches were restricted to microwave frequencies, this framework extends the paradigm into the visible and near-infrared spectrum. Future applications will target high-Q resonators, nonlinear optical platforms, and three-dimensional metastructures. Beyond photonics, researchers project that reality-infused deep learning will become a standard architecture for materials science, electronic engineering, and quantum system development.