Topology-Driven Automatic Discovery of Optimal Microwave Photonic Architectures

When photonic platforms are used to handle microwave signals, the system inherits wide optical bandwidths, minimal transmission loss, and a welcome immunity to electromagnetic noise. These benefits have become increasingly important as radar, wireless communication, and data-processing systems push toward higher frequencies and larger dynamic ranges. Recently, improvements in integrated photonic fabrication especially silicon photonics and a range of heterogeneous material stacks have broadened what designers can actually implement on chip. Functions such as carefully shaped microwave filters, chip-level differentiators, and compact arbitrary-waveform generators are now realistic building blocks for full systems. However, despite the sophistication of these platforms, the way microwave photonic systems are designed remains surprisingly traditional. Most researchers approach a new function by reaching for well-known components—FIR filters for generality or IIR structures based on microring resonators when high-Q shaping is required. Both choices come with limitations that everyone in the field is aware of: FIR filters can approximate almost anything if one is willing to accept a long tap chain, and IIR filters stay compact but inevitably restrict the frequency response to a narrow family of shapes. When the two are hybridized, the design space widens a bit, though not nearly enough to explore all the architectures that are mathematically possible. As a result, many potentially superior configurations most likely sit untouched, simply because they do not resemble the circuit motifs that the field has grown comfortable with. The situation becomes more complicated as photonic chips incorporate larger routing networks, more elaborate cascades of resonators, and tunable elements that change the system’s operating point. Once these features are available, the number of feasible architectures expands extremely quickly—so quickly that intuition offers little guidance about where the “good” ones might lie. Attempting to sort through these possibilities manually becomes unrealistic, and even experienced designers cannot easily tell whether a proposed configuration is anywhere near optimal. In practice, this creates a widening gap between what integrated photonics enables and what conventional microwave-photonic design methods can systematically explore.

 To this account, new research paper published in Optics Express and conducted by Mr. Bo Li, Prof. Shaofu Xu, Ms. Ting Lyu, Mr. Ruicheng Qiu, Ms. Yixi Liu, and Prof. Weiwen Zou from the Department of Electronic Engineering at Shanghai Jiao Tong University, researchers developed an automated design framework that represents microwave photonic systems as complete topological graphs and searches this space systematically. They combined exhaustive topology generation with a hybrid genetic-algorithm and L-BFGS parameter search to identify architectures that match target responses with minimal system cost. The method consistently uncovers non-intuitive designs that outperform conventional FIR and IIR configurations. Across differentiators, waveform generators, Hilbert transformers, and pulse-compression systems, the framework delivers architectures that are simultaneously more efficient and more accurate than existing hand-designed counterparts.

The authors defined a structured way to represent any microwave photonic system as a graph. Each processing block—whether an FIR filter or one of the IIR variants—is treated as a node with tunable parameters. Connections between nodes form the edges, encoding whether units operate in series or in parallel. Once this abstraction is established, the team develops an exhaustive graph-generation routine that enumerates all feasible directed acyclic graphs constructed from a chosen number of basic units. The algorithm then converts each graph into an associated analytical expression using depth-first search, ensuring that the full diversity of transfer functions emerging from these combinations is preserved. Redundant expressions are removed to avoid unnecessary computation during later stages.  The researchers then turned to parameter optimization. Here, the task is to tune the coefficients, delays, and loop gains of the processing units so that the resulting system response matches a target function. Depending on the application, the target may be an ideal mathematical operator—such as a temporal differentiator—or an output waveform directly, as in the case of arbitrary waveform generation. To balance accuracy against hardware complexity, the loss function includes both a mean-square-error term and a cost term proportional to the number of tunable parameters inside the architecture. Minimizing this loss requires navigating a high-dimensional, non-convex space, so the team employs a hybrid scheme: a genetic algorithm performs a global search, and the L-BFGS method refines promising candidates to convergence.  They applied the workflow to several benchmark functionalities and found for the second-order temporal differentiator, the automated search not only rediscovers the known cascaded IIR configuration but also produces a non-intuitive alternative that achieves lower error with fewer parameters. A similar story unfolds with the Hilbert transformer. The framework identifies an architecture capable of matching the desired π/2 phase shift while reducing amplitude ripple relative to the conventional 12-tap FIR design. In both cases, the system’s output for Gaussian and sinusoidal test signals demonstrates that these automatically derived architectures behave exactly as required. Furthermore, the authors reported for sawtooth, triangular, and square waveform generation, the search discovers compact structures formed from multiple FIR filters arranged in mixed series–parallel layouts. These achieve lower mean-square errors than the standard single-filter approach and do so with reduced tap counts. Similar gains are observed in linear-frequency-modulated pulse compression, where a newly found architecture significantly improves peak side-lobe ratio while using less than half the parameter budget of the conventional 20-tap FIR implementation. Even in millimeter-wave signal generation, the algorithm selects a single bandstop IIR filter as the optimal spectral shaper, cutting system cost and also preserving side-mode suppression. The team also conducted a robustness study which further showed that these architectures remain resilient under typical fabrication errors, and retain performance advantages over their conventional counterparts.

In conclusion, the new work by Shanghai Jiao Tong University scientists demonstrated that microwave photonic architectures can be discovered automatically, the authors open the door to a fundamentally different design culture—one in which intuition serves as a starting point but no longer dictates the boundaries of what is possible. Their results show, often dramatically, that optimal solutions do not always resemble familiar structures. In some cases, the best architecture contains unexpected combinations of parallel and series paths; in others, it relies on a deliberately minimal set of components whose utility only becomes apparent once the loss landscape is mapped numerically. The new automated framework evaluates complexity explicitly through its cost term, rewarding architectures that achieve performance targets with fewer adjustable parameters. This naturally favors solutions that remain compact and physically implementable on integrated platforms and with the photonic chips continue to grow in scale and functionality, this cost-aware optimization will become critical, especially when designers must accommodate tight power budgets, limited chip area, or stringent thermal constraints.

Scalability is an important advantage of the new design and the graph-based representation provides a direct pathway toward more sophisticated systems, including multi-input multi-output configurations or nonlinear photonic processors. Although the study focused on linear, single-input single-output systems, the framework itself is not inherently limited to these cases. Adjustments to node types and loss definitions could, in principle, enable automatic discovery of architectures for optical neural networks, multi-channel sensing systems, or photonic compute blocks. The ability to explore these expanded design spaces algorithmically could accelerate progress across a spectrum of emerging technologies that are beginning to rely on photonic signal processing. Another advantage is the framework’s robustness to fabrication imperfections. The authors show that even when amplitude and phase errors of several percent are introduced, the automatically discovered architectures maintain performance advantages over conventional solutions which suggests that the framework produce mathematically optimal designs and also as important identifies physically resilient ones. In sum, the authors create a toolbox capable of uncovering architectures that combine efficiency, performance, and implementation. As microwave photonics continues expanding into areas such as high-capacity wireless communication, cognitive sensing, and photonic computing, the impact of such a methodology is likely to grow.