Chinese Scientists Build Programmable 3D Photonic Neural Network Inside Glass

A team of researchers from Huazhong University of Science and Technology (HUST) and Shanghai Jiao Tong University has demonstrated a remarkable new paradigm for photonic computing: programmable three-dimensional neural networks written directly inside glass using ultrafast lasers. The work, published under the architecture name LAMP — short for Lantern-shaped Adaptive Multifunctional Photonic computing — could point the way toward high-dimensional AI inference and signal processing that leaves electronic chips in the dust. Photonic computing replaces electrons with photons as the information carrier. Light propagates faster than electrical signals, generates far less heat, and — crucially — supports parallel data encoding across multiple physical dimensions such as wavelength, phase, and polarization simultaneously. This makes photonics a natural fit for the matrix-multiplication-heavy workloads that dominate modern artificial intelligence. The bottleneck has always been programmability: most photonic chips are fabricated as static circuits, hardwired for one task and then frozen in silicon. The LAMP architecture shatters that limitation. Using femtosecond laser direct writing — ultrashort pulses that modify the refractive index of glass at microscopic scale — the team etched complex three-dimensional waveguide structures directly into a glass substrate. These waveguide lattices form a photonic neural network whose connections can be reconfigured optically, allowing the same physical chip to perform different inference tasks without any changes to the underlying hardware. Why glass? The material offers exceptional optical clarity, thermal stability, and mechanical rigidity — all critical for guiding photons with minimal loss over complex 3D paths. And because the writing process is subtractive only during fabrication (the laser modifies the glass internally, leaving the surface pristine), the resulting device is a monolithic block of transparent material with no moving parts and no layered assembly. The implications are significant. High-dimensional data — hyperdimensional vectors, multi-channel sensor fusion, spectral imaging — can be processed directly on-chip without being flattened into 1D representations as conventional electronics require. The researchers, led by Prof. Zhang Xinliang and Prof. Dong Jianji at HUST alongside Prof. Tang Hao at SJTU, see applications in AI inference acceleration, real-time optical signal processing, and next-generation optical neural networks that could operate at speeds and energy efficiencies beyond the reach of CMOS. While commercial deployment remains years away — femtosecond laser fabrication is slow and expensive at scale — the LAMP architecture proves that programmable, three-dimensional photonic computing inside glass is not only possible but practical. It is the kind of foundational breakthrough that rewires the roadmap for what an AI chip can look like.