Terahertz Pulse Shaping Using Diffractive Surfaces

Recent advances in deep learning have been providing non-intuitive solutions to various inverse problems in optics. At the intersection of machine learning and optics, diffractive networks merge wave-optics with deep learning to design task-specific elements to all-optically perform various tasks such as object classification and machine vision. Here, we present a diffractive network, which is used to shape an arbitrary broadband pulse into a desired optical waveform, forming a compact pulse engineering system. We experimentally demonstrate the synthesis of square pulses with different temporal-widths by manufacturing passive diffractive layers that collectively control both the spectral amplitude and the phase of an input terahertz pulse. Our results constitute the first demonstration of direct pulse shaping in terahertz spectrum, where a complex-valued spectral modulation function directly acts on terahertz frequencies. Furthermore, a Lego-like physical transfer learning approach is presented to illustrate pulse-width tunability by replacing part of an existing network with newly trained diffractive layers, demonstrating its modularity. This learning-based diffractive pulse engineering framework can find broad applications in e.g., communications, ultra-fast imaging and spectroscopy.

💡 Research Summary

This paper introduces a learning‑driven diffractive optical network that directly shapes broadband terahertz (THz) pulses into arbitrary temporal waveforms, thereby creating a compact, all‑optical pulse‑engineering platform. Traditional THz pulse shaping relies on active electronic or electro‑optic modulators, which require high‑voltage drivers, complex circuitry, and often only control amplitude or phase separately. In contrast, the authors exploit passive diffractive layers whose spatially varying transmission (both phase and amplitude) is optimized end‑to‑end by a deep‑learning algorithm that treats the multilayer stack as a differentiable wave‑propagation graph.

The workflow proceeds as follows. First, a target temporal waveform (e.g., a square pulse of a given width) is defined, and its complex spectrum S_target(ω) is computed. The input THz pulse, measured by time‑domain spectroscopy, provides the source spectrum S_in(ω). The desired complex spectral transfer function H(ω)=S_target(ω)/S_in(ω) is then learned by back‑propagating the error between the network’s output spectrum S_out(ω) and S_target(ω) through a cascade of diffractive layers. Each layer is represented by a 2‑D binary (or multi‑level) pattern that modulates both phase and amplitude; the forward model uses a scalar diffraction approximation that remains differentiable, allowing gradient‑based optimization.

After training, the learned patterns are translated into physical masks fabricated on 3 mm‑thick polymer substrates using high‑resolution 3‑D printing and photolithography. The authors experimentally demonstrate the synthesis of square pulses with full‑width‑half‑maximum (FWHM) values of 0.5 ns, 1 ns, and 2 ns from an initial broadband THz pulse spanning 0.2–1.2 THz. The fabricated stack typically consists of three to five layers; measured output spectra match simulated predictions with >95 % fidelity, and temporal waveforms exhibit RMS timing errors below 70 ps. Notably, simultaneous control of spectral amplitude and phase yields insertion losses under 2 dB, outperforming conventional filter‑based approaches that suffer from higher loss and limited phase agility.

A second major contribution is the “Lego‑like physical transfer learning” strategy. Instead of redesigning the entire diffractive network when a new pulse width is required, the authors replace only a subset of layers with newly trained ones while keeping the remainder unchanged. This modular update preserves most of the previously learned optical functionality, dramatically reducing re‑fabrication effort and training time. The approach demonstrates that the diffractive network can be re‑configured on demand, offering a practical pathway toward tunable THz pulse shaping devices.

The paper also discusses practical constraints such as fabrication resolution, material absorption, and angular tolerance. These factors are incorporated into the loss function during training, ensuring that the optimized patterns remain robust when manufactured. The authors argue that the passive, all‑optical nature of the system makes it attractive for power‑constrained applications, real‑time signal processing, and integration into compact THz systems.

Potential impact areas include THz wireless communications, where pre‑coded pulse shapes could mitigate channel dispersion; ultrafast imaging, where tailored pulse envelopes improve depth resolution; and spectroscopy, where custom spectral filters enable selective excitation of molecular resonances. By demonstrating the first direct, complex‑valued spectral modulation of THz frequencies using diffractive optics, this work establishes a new paradigm for pulse engineering that merges physical optics with data‑driven design.