⚡ electrical engineering

Fabrication-constrained tandem inverse design of photonic crystal fiber SPR sensors with uncertainty-aware Pareto screening

This paper presents a fabrication-constrained tandem inverse-design framework for photonic crystal fiber SPR sensors that utilizes differentiable regularization and multi-metric Pareto screening to explicitly balance geometric feasibility, out-of-distribution detection, and target accuracy while addressing the non-uniqueness of inverse solutions.

Original authors: Md Shahanur Islam Shagor

Published 2026-09-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Md Shahanur Islam Shagor

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a world where light is trapped inside a glass thread, guided not by a solid core but by a delicate, honeycomb-like pattern of tiny air holes running along its length. This is the realm of photonic crystal fibers, a type of optical cable that engineers can shape with extreme precision to control how light behaves. When these fibers are coated with a thin layer of metal, they can interact with the surrounding environment in a unique way. If a liquid flows through a channel in the fiber, the light inside changes its color or intensity in response to the liquid's properties. This phenomenon, known as surface plasmon resonance, turns the fiber into an incredibly sensitive sensor, capable of detecting minute changes in chemicals or biological agents. The challenge for scientists has always been the reverse problem: instead of building a fiber and seeing what it does, how do you design the perfect fiber from scratch to detect a specific target? It is like trying to guess the exact shape of a key that will open a specific lock, knowing that many different keys might fit, but only one is strong enough to be manufactured without breaking.

In a recent study, a researcher at Voronezh State University of Forestry and Technologies tackled this design challenge by creating a new computer method that respects the physical limits of manufacturing. The work focuses on a specific type of sensor where the geometry of the fiber—the size of the holes, the thickness of the metal, and the spacing of the pattern—determines its performance. The researcher developed a two-part artificial intelligence system. The first part acts as a fast simulator, learning to predict how a specific fiber shape will react to different liquids. The second part works backward, taking a desired performance goal and generating a list of possible fiber shapes that could achieve it. Crucially, this system does not just guess; it is trained to understand that some shapes, while mathematically perfect on a screen, are impossible to build in a real lab. The method explicitly penalizes designs that would require holes to overlap or metal layers to be thinner than physically possible, ensuring that every suggestion is a candidate that could actually be made.

The study revealed that finding a single perfect shape is often impossible because different designs can produce the same result. To handle this, the system generates a population of candidate shapes rather than forcing a single answer. It then filters these candidates through a rigorous screening process. It checks if the design falls within the safe zone of known data, ensuring the computer is not making wild guesses about shapes it has never seen. It also ranks the candidates based on a balance of competing goals: how well they meet the sensing target, how close they are to being manufacturable, and how much uncertainty exists in the prediction. This approach allows the system to present a set of trade-offs, showing the user that a slightly less sensitive sensor might be much easier to build, or that a highly sensitive one might carry a higher risk of failure.

One of the most significant findings was that the way data is organized matters just as much as the algorithms used. The researcher discovered that if the computer training data is split randomly, it can lead to memorization of the same fiber shape under different conditions, leading to false confidence in its predictions. By grouping all data points belonging to the same fiber shape together and keeping them in either the training or testing phase, the system learned to generalize correctly. The results showed that when the system was forced to respect manufacturing limits, the number of valid, buildable designs it produced increased from about 79 percent to 88 percent. However, this safety came with a cost: the designs were slightly less perfect at hitting the exact numerical target, dropping in accuracy from a score of 0.830 to 0.786. This trade-off is a fundamental reality of engineering; making a design safer and more realistic often means sacrificing a small amount of theoretical perfection.

The study also highlighted a subtle but important flaw in how sensor performance is usually measured. The researchers found that while the computer could easily predict the specific color where the sensor reacts, it struggled to predict the sensitivity—the rate at which that reaction changes. This is because sensitivity is not a single point on a graph but a relationship derived from how the sensor behaves across a range of conditions. The study suggests that future designs should focus on predicting the raw behavior of the sensor first and calculating the sensitivity afterward, rather than trying to predict the sensitivity directly. This distinction prevents the computer from learning noise and errors that arise from trying to estimate a rate of change from limited data points.

Ultimately, this work does not claim to have built a new, record-breaking sensor. Instead, it provides a robust, mathematically transparent workflow for designing one. The numbers and results presented are derived from a synthetic computer simulation designed to test the logic of the method, not to claim a physical breakthrough. The value lies in the framework itself: a reproducible path that takes a performance goal, generates feasible designs, screens them for risk and manufacturability, and hands a specific, verified geometry to a separate, high-precision physics simulator for final confirmation. By separating the software logic from the physical evidence, the researcher has created a tool that can guide engineers toward designs that are not only smart on paper but also viable in the real world, bridging the gap between digital optimization and physical fabrication.

Technical Summary: Fabrication-constrained tandem inverse design of photonic crystal fiber SPR sensors with uncertainty-aware Pareto screening

Problem Statement
The inverse design of Photonic Crystal Fiber Surface Plasmon Resonance (PCF-SPR) sensors faces three primary challenges:

Non-uniqueness: Different geometric configurations can produce identical spectral responses, making a single-valued inverse mapping mathematically ill-posed.
Fabrication Constraints: Numerically optimal solutions often violate physical manufacturing bounds (e.g., air-hole diameters exceeding pitch or negative thicknesses).
Data Leakage and Metric Definition: Sensitivity and Figure of Merit (FOM) are derived quantities calculated from refractive-index sweeps of a fixed geometry. Standard row-wise data splitting in machine learning can leak geometric information between training and validation sets, inflating generalization metrics. Furthermore, treating sensitivity as an independent row-wise observable ignores its relational nature.

Methodology
The paper proposes the CyberPhotonics-SPR framework, a fabrication-constrained tandem inverse-design workflow that integrates uncertainty quantification and multi-objective screening.

Data Handling: The methodology enforces a "grouped split" protocol where all refractive-index points belonging to a single base geometry remain in the same data partition (training, validation, or calibration). This prevents leakage of geometric trends.
Tandem Architecture:
- Forward Surrogate: A Multi-Layer Perceptron (MLP) maps four geometric variables (pitch, normalized air-hole diameter, gold thickness, channel radius) and analyte refractive index to three targets: resonance wavelength, wavelength sensitivity, and FOM.
- Inverse Generator: A conditioned generator produces a population of candidate geometries for a requested sensing target. It utilizes a latent vector to address the one-to-many nature of the inverse problem, allowing the generation of multiple valid geometries for a single target.
Training and Regularization:
- The forward model is trained first and then frozen.
- The inverse generator is optimized to minimize the mismatch between the forward-predicted response and the target, while simultaneously minimizing differentiable fabrication penalties. These penalties enforce physical bounds (e.g., $d < \Lambda$ ) and prevent invalid geometries before post-hoc projection.
Uncertainty and Screening:
- Conformal Prediction: Computes residual intervals on a held-out set to provide distribution-free reliability statements for predictions.
- OOD Detection: Uses Mahalanobis distance to screen candidates that lie outside the training distribution (extrapolation risk).
- Pareto Ranking: Candidates are ranked using a multi-objective vector comprising target error, fabrication projection distance, and OOD distance. This preserves trade-offs rather than collapsing them into a single loss.
Validation Protocol: An independent "fixed-geometry" verification path recomputes sensitivity and FOM by fitting a line through the complete refractive-index sweep of a candidate geometry, rather than relying on row-wise finite-difference labels used during training.

Key Results
The study utilizes a software-consistency benchmark involving 100 synthetic geometries and five refractive-index points per geometry. The results are presented as validation of the computational pipeline, not as claims of physical sensor performance.

Target Asymmetry: Baseline models (Ridge, Random Forest, MLP) demonstrated that resonance wavelength is easily learned ( $R^2 \approx 1.0$ ), whereas row-wise local sensitivity is poorly modeled ( $R^2 < 0$ ). This confirms that sensitivity is a relational quantity that cannot be treated as an independent row-wise target without introducing noise.
Feasibility-Accuracy Trade-off: A three-seed ablation study compared an unconstrained inverse generator against a fabrication-regularized one:
- Raw Geometry Validity: Increased from 79.0% (unconstrained) to 88.3% (regularized).
- Inverse Target Satisfaction ( $R^2$ ): Decreased from 0.830 to 0.786.
- This explicitly demonstrates that fabrication constraints restrict the solution space, reducing the ability to perfectly match the surrogate target but increasing the physical feasibility of the generated designs.
Synthetic Consistency: The fixed-geometry validation path successfully recovered the linear refractive-index trend embedded in the synthetic generator (Mean sensitivity $\approx 820$ nm/RIU, $R^2_{linearity} \approx 0.99999$ ), confirming the internal consistency of the metric extraction and grouping logic.

Significance and Claims
The paper explicitly positions its contribution as a mathematical and computational integration rather than a new physical sensor discovery.

Methodological Novelty: The work integrates inverse generation, feasibility constraints, uncertainty quantification (conformal prediction), distribution shift detection (Mahalanobis), and Pareto ranking into a single reproducible workflow. It addresses the "non-unique" inverse problem by generating a population of candidates rather than a single point estimate.
Evidence Discipline: The author emphasizes a strict separation between software-consistency evidence (validating code paths and logic) and physics-ready evidence (electromagnetic simulation or experimental data). The reported numerical values (sensitivity, FOM) are synthetic benchmarks used to verify the methodology, not claims of device performance.
Traceability: The framework provides a "mathematically traceable route" from target specification to risk-aware candidate selection. It allows designers to inspect why a candidate was rejected (e.g., high OOD distance vs. high target error) rather than relying on a black-box score.
Future Path: The paper concludes that the next necessary step is to apply this exact pipeline to verified Finite Element Method (FEM) or COMSOL models on disjoint train/calibration/test sets to generate physical sensor-performance claims.

In summary, the paper presents a robust, uncertainty-aware inverse design framework that prioritizes fabrication feasibility and rigorous validation protocols over raw predictive accuracy, offering a structured approach to navigating the complex design space of PCF-SPR sensors.

Technical Summary: Fabrication-constrained tandem inverse design of photonic crystal fiber SPR sensors with uncertainty-aware Pareto screening

More like this