Advancing Photonics Through Data-Driven Innovation at SUTD
Singapore University of Technology and Design continues to push boundaries in engineering and design education by fostering research that bridges theoretical models with practical fabrication realities. A recent breakthrough from the institution highlights how artificial intelligence can transform the design of optical nanostructures by learning directly from experimental outcomes rather than relying solely on idealized computer simulations.
This development comes at a time when Singapore's higher education sector emphasizes interdisciplinary approaches, combining photonics, machine learning, and materials science to address real-world challenges in sensing, displays, and communications. The work exemplifies the kind of applied research that prepares graduates for careers in emerging technologies while strengthening the nation's position as a hub for advanced manufacturing and innovation.
Understanding Optical Fourier Surfaces and Their Design Challenges
Optical Fourier surfaces are specialized nanostructured gratings engineered to precisely control light behavior at the nanoscale. These surfaces redistribute incoming light into specific directions and wavelengths, making them valuable for applications such as compact spectrometers, augmented-reality displays, advanced sensors, and next-generation photonic devices. In Singapore's context, where precision engineering supports industries from semiconductors to biomedical imaging, such technologies hold significant promise for both academic research and industrial translation.
Traditional design methods for these surfaces depend heavily on computer simulations like finite-difference time-domain modeling. These simulations typically assume perfect geometries, smooth surfaces, and single-angle illumination, conditions that rarely match the complexities of actual fabricated devices. Fabrication processes introduce roughness, structural asymmetries, and material variations, while real-world measurements include noise and angular distributions rather than idealized single-incident angles. This mismatch often leads to designs that perform differently in practice, requiring multiple costly fabrication and testing cycles.
Associate Professor Dong Zhaogang from SUTD has noted that incorporating the angle of incoming light as a design parameter adds a powerful degree of freedom, allowing a single structure to perform multiple functions without physical changes. However, simulations at oblique angles tend to be computationally unstable and expensive, limiting practical use until now.
The ExpForm AI Framework: A Reality-Infused Approach
Researchers at SUTD, in collaboration with colleagues from Xiamen University and Hefei University of Technology, developed ExpForm, a transformer-based neural network designed to overcome these limitations. Unlike conventional models trained on simulated data, ExpForm learns directly from experimental measurements of fabricated nanostructures. This reality-infused training enables the AI to inherently account for fabrication imperfections, measurement noise, and real angular variations.
The framework uses a high-throughput, angle-resolved spectroscopy system to capture broadband reflectance spectra across wide ranges of incident and azimuthal angles. Data collection from four quasi-optical Fourier surface samples, fabricated via nanoimprint lithography, yielded over 25,000 spectral instances. These real spectra, complete with actual imperfections, formed the training dataset.
ExpForm operates bidirectionally. The forward network predicts optical spectra from structural and angular inputs in real time. The inverse network works backward from a desired spectral response to determine the required structural dimensions and illumination angles. Together, they create an end-to-end design tool that accelerates iteration and reduces reliance on repeated fabrication trials.
Performance Metrics and Comparative Advantages
Benchmarking against traditional finite-difference time-domain simulations revealed impressive results. ExpForm achieved 99.79 percent consistency with experimental measurements while providing approximately 900 times faster spectral evaluation. In test cases involving oblique and azimuthal incidence, conventional simulations often missed key features such as wavelength positions, resonance shapes, and quality factors, whereas the AI model captured them accurately.
This speed and accuracy translate to practical benefits for researchers and engineers. Design cycles shrink from hours or days to seconds, enabling rapid prototyping and exploration of complex spectral responses. The approach supports on-demand generation of narrowband resonances, high-reflectance profiles, and dual-band resonances simply by adjusting incident angles, without needing new fabricated structures each time.
The team has made the full experimental training dataset publicly available to support broader research and fair benchmarking of neural network models in photonics. This openness aligns with Singapore's emphasis on collaborative innovation within its higher education ecosystem.
Photo by Luke Jones on Unsplash
Broader Implications for Singapore's Higher Education and Research Landscape
The ExpForm project underscores SUTD's role in Singapore's higher education sector as a leader in design-centric, AI-integrated engineering education. By training students and researchers in data-informed design paradigms, the university prepares graduates for roles in photonics, materials science, electronics, and quantum technologies. Such skills are increasingly sought after in Singapore's growing semiconductor and advanced manufacturing sectors.
Collaborations spanning Singapore and China institutions also reflect the international outlook encouraged in local universities. This work extends experiment-driven AI approaches from microwave frequencies into the visible and near-infrared ranges, opening new avenues for compact, multifunctional optical devices relevant to Singapore's smart nation initiatives and biomedical applications.
Future extensions could include high-Q resonators, nonlinear optical platforms, and three-dimensional metastructures, as well as dielectric devices leveraging bound states in the continuum. The paradigm shift toward AI as a co-designer rather than just a computational tool has potential applications beyond photonics in fields like materials science and quantum devices.
Challenges Addressed and Remaining Considerations
While ExpForm demonstrates clear advantages, scaling such models requires substantial experimental datasets and robust spectroscopy infrastructure. Singapore universities like SUTD invest in cleanroom facilities and advanced characterization tools to support this type of research, providing students with hands-on experience in both fabrication and data-driven analysis.
Ensuring model generalizability across different materials and fabrication methods remains an ongoing area of development. The public release of the dataset helps address this by enabling community-wide validation and improvement.
Integration into educational curricula could involve modules on AI for photonics design, giving PhD-track students and researchers practical exposure to inverse design tools that complement traditional simulation methods.
Future Outlook and Opportunities for Academics and Institutions
As Singapore continues to position itself as a global leader in photonics and AI, projects like this from SUTD highlight opportunities for faculty and researchers to contribute to high-impact publications and industry partnerships. The emphasis on experimental grounding aligns with national priorities for translational research that moves from lab to application efficiently.
Academic job seekers interested in Singapore's higher education sector may find roles at institutions prioritizing interdisciplinary AI-photonics research particularly rewarding. Universities are actively recruiting talent to expand capabilities in machine learning for physical sciences and advanced optical technologies.
Broader adoption could accelerate innovation in areas such as augmented reality, sensing, and sustainable energy technologies, areas where Singapore seeks to maintain competitive edges.
Supporting Singapore's Research Ecosystem Through Open Science
The decision to release the experimental dataset publicly supports open science principles increasingly valued in Singapore's academic community. This practice lowers barriers for other groups and promotes reproducible research, consistent with efforts by local institutions to enhance research integrity and collaboration.
Faculty at SUTD and similar universities often emphasize such practices in mentoring graduate students, preparing them for careers where data sharing and interdisciplinary teamwork are essential.
Photo by Luke Jones on Unsplash
Conclusion: A Step Forward in Reality-Aware AI Design
The SUTD-led development of ExpForm represents a meaningful advance in using AI for optical nanostructure design by embracing real-world complexities. By training on experimental data, the model delivers faster, more reliable results than simulation-only approaches, with clear benefits for research efficiency and device performance.
This work reinforces Singapore's higher education strengths in fostering innovative, applied research that prepares the next generation of engineers and scientists. As the field evolves, continued investment in such projects will support both academic excellence and economic impact in key technology sectors.