🔬 physics

Differentiable surrogate modeling for strain-engineered exciton localization in hybrid TMD-photonic nanostructures

This paper presents a differentiable surrogate modeling framework that combines a finite-difference strain solver with a trained convolutional neural network to enable rapid, gradient-based inverse design of photonic substrates for deterministic exciton localization in hybrid TMD nanostructures.

Original authors: Georgii Bulavin, Alexander Shorokhov, Andrey Fedyanin

Published 2026-09-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Georgii Bulavin, Alexander Shorokhov, Andrey Fedyanin

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 not just a wave, but a stream of individual particles, each carrying a single piece of information. This is the realm of quantum photonics, a field dedicated to building computers and communication networks that operate on the fundamental rules of the universe. To make these machines work, scientists need a reliable way to create sources that emit these single particles of light on demand. One of the most promising materials for this task is a class of ultra-thin crystals known as transition metal dichalcogenides. These materials are only one atom thick, yet they possess remarkable optical properties that make them ideal for trapping and releasing light. However, a major hurdle has always stood in the way of using them effectively: these crystals are incredibly sensitive to their physical environment. When placed on a surface, even the slightest bump or curve in that surface can change how the crystal behaves, often in unpredictable ways. For years, researchers have struggled to control this sensitivity, unable to predict exactly where a single particle of light would appear or how to guide it to a specific spot.

A team of researchers at Lomonosov Moscow State University has now developed a new method to solve this problem, turning a source of unpredictability into a precise tool for design. Their work focuses on a technique called strain engineering, which involves deliberately bending or stretching a material to change its properties. In the case of these atom-thin crystals, stretching them creates a landscape of energy that acts like a funnel, guiding the light-emitting particles toward specific locations. The challenge has been that calculating exactly how a crystal will bend over a complex, patterned surface is a slow and difficult mathematical task, often requiring hours of computer time for a single design. This made it nearly impossible to iterate quickly and find the perfect shape needed to trap light exactly where engineers wanted it.

To overcome this bottleneck, the researchers built a sophisticated computer model that mimics the physical behavior of the crystal. They started by creating a digital solver based on the physics of thin plates, a well-established theory that describes how flat materials bend under pressure. Using this solver, they simulated thousands of different scenarios, placing the crystal over various patterns of pillars and holes to see how the strain would distribute across the surface. They generated a massive dataset of over six thousand simulations, each mapping a unique substrate shape to the resulting strain pattern in the crystal. This data served as the training ground for a new type of artificial intelligence, specifically a neural network designed to recognize patterns in images.

The result is a digital surrogate, a fast and highly accurate AI model that can predict the strain landscape in a fraction of a second. While the original physics-based solver took over five seconds to compute a single strain map, the new AI model does the same job in less than three milliseconds. This represents a speedup of nearly two thousand times, transforming a process that was once too slow for practical design into one that can happen in real time. More importantly, the AI model is differentiable, meaning it can not only predict the outcome but also calculate how to change the input to achieve a desired result. This capability allows the researchers to work backward from a goal. Instead of guessing a shape and seeing what happens, they can now specify exactly where they want the light-emitting particles to gather, and the system will automatically design the substrate pattern required to create that specific strain field.

The researchers tested this approach by designing a substrate configuration intended to concentrate the strain at the center of a waveguide, a channel used to direct light. By feeding this goal into their optimization loop, the system generated a new, complex pattern of pillars and voids that had never been seen before. When they analyzed the results, they found that the optimized design successfully funneled the light-emitting particles toward the target area, increasing the concentration of these particles by thirty-one percent compared to a standard, unmodified surface. The design also improved the alignment between the particles and the light-carrying channel by twenty-one percent, a significant gain for the efficiency of the device. Crucially, the researchers demonstrated that this improvement was achieved without altering the optical properties of the waveguide itself; they simply changed the mechanical environment beneath the crystal.

The study confirms that this method works not just for the specific materials they tested, but offers a general framework that can be applied to other thin-film systems. The team has made their entire pipeline, including the code and the data, available to the public, allowing other scientists to adapt the tool for different materials or applications. While the current model relies on simulations and assumes ideal conditions, the researchers note that the approach provides a reliable guide for designing real-world devices. By decoupling the mechanical design from the optical design, this work opens a direct path toward creating integrated quantum photonic platforms where the placement of light sources is no longer a matter of chance, but a matter of precise, deterministic engineering. The ability to co-design the mechanical and optical responses of these hybrid structures marks a significant step forward in the development of scalable quantum technologies.

Technical Summary: Differentiable Surrogate Modeling for Strain-Engineered Exciton Localization

Problem Statement
Transition metal dichalcogenide (TMD) monolayers are promising platforms for integrated quantum photonics, particularly as on-chip single-photon emitters. However, the formation of these emitters is highly sensitive to local mechanical strain, which creates a potential landscape that drives exciton drift and localization. While strain engineering is a known strategy for positioning emitters, a quantitative framework linking the photonic substrate configuration to the resulting strain landscape in TMD films has been lacking. Current design workflows typically optimize optical figures of merit (e.g., transmission, Q-factor) without accounting for the mechanical response of the overlying 2D material. This disconnect leads to substantial device-to-device variability and prevents deterministic emitter placement in hybrid silicon nitride (SiN)–TMD nanostructures. Furthermore, direct numerical simulation of the mechanical forward problem (solving the Kirchhoff–Love plate equation) is computationally expensive, rendering it impractical for iterative inverse design or topology optimization, which require thousands of forward evaluations.

Methodology
The authors propose a three-stage computational framework to bridge the gap between substrate topology and exciton localization:

Physics-Based Forward Solver: A finite-difference method (FDM) based on Kirchhoff–Love plate theory is implemented to compute spatially resolved in-plane strain maps. The model discretizes the plate equation on a uniform $256 \times 256$ grid ( $12 \times 12$ µm domain) using a 13-point stencil. The resulting strain field $\varepsilon(x, y)$ is coupled to the excitonic potential via the deformation potential relation $\Delta E_g = -\alpha \varepsilon$ . This potential drives a drift-diffusion equation to predict steady-state exciton density and localization sites. The solver was validated against experimental Raman data from literature, showing a mean absolute error below 13%.
Differentiable Surrogate Modeling: To overcome the computational bottleneck of the FDM solver (approx. 5060 ms per sample), the authors trained a convolutional neural network (CNN) as a surrogate. The network follows a U-Net encoder-decoder architecture with skip connections. It takes a three-channel input: a binary support map of the substrate and a two-channel coordinate grid. Trained on a dataset of 6,300 FDM simulations covering 21 distinct configuration classes, the surrogate learns to map substrate geometry directly to strain fields. The training objective combines Mean Squared Error (MSE) with a gradient penalty term to ensure accurate prediction of strain gradients at feature boundaries.
Gradient-Based Inverse Design: The trained surrogate is embedded into a topology optimization loop. The substrate configuration is parameterized in logit space, and a composite loss function is minimized to maximize the strain contrast between a target zone and the background. The optimization employs a three-phase curriculum schedule: (I) establishing coarse strain localization, (II) enforcing bilateral symmetry, and (III) applying a fill-factor penalty. The differentiable nature of the surrogate allows for efficient gradient-based updates to the substrate topology.

Key Results

Surrogate Accuracy and Speed: The surrogate network achieves a coefficient of determination $R^2 = 0.9945$ on the held-out test set, capturing 99.5% of the variance in the FDM strain fields. In terms of computational efficiency, the surrogate provides a 1,868× speedup over the direct FDM solver at the training resolution ( $256 \times 256$ ), reducing inference time from ~5 seconds to 2.71 ms. On a larger $1024 \times 1024$ grid, the speedup increases to 5,082×.
Generalization: The model demonstrates strong out-of-distribution generalization, achieving $R^2 = 0.983$ on a resonant waveguide configuration not present in the training data.
Inverse Design Performance: Using the surrogate in a topology optimization loop, the authors designed a substrate configuration that concentrates strain at a specific target zone within a waveguide. Compared to a bare waveguide baseline, the optimized design increased the peak exciton concentration in the target zone by 31% and improved the overlap integral between the exciton density and the guided TM mode electric field by 21%. The resulting optimized geometry features physically interpretable, symmetric pillar-like protrusions that focus substrate curvature toward the waveguide center.

Significance and Claims
The paper claims to establish a general, transferable methodology for the deterministic placement of quantum emitters in hybrid photonic devices. By decoupling the mechanical strain engineering from the optical design, the framework allows strain optimization to be applied as a post-design step to existing photonic layouts without compromising their optical performance. The authors emphasize that this approach provides the computational foundation for "multiphysics co-design," where optical and mechanical responses are engineered simultaneously. While the current work focuses on the WSe $_2$ –SiN system, the authors note that the framework is material-agnostic and can be transferred to other thin-film systems (e.g., hBN, MoS $_2$ ) by updating material parameters. The study concludes that this differentiable surrogate approach opens a direct path toward the inverse design of integrated quantum photonic platforms, addressing a critical limitation in current device fabrication where emitter placement remains stochastic.

Technical Summary: Differentiable Surrogate Modeling for Strain-Engineered Exciton Localization

More like this