Exploring polymer classification with a hybrid single-photon quantum approach

TL;DR

This work addresses polymer gap classification by integrating a classical deep neural network featurizer with a photonic single-photon variational quantum classifier. The hybrid classical–quantum pipeline encodes compact 4D polymer features into a linear-optical circuit trained via seesaw optimization, and its performance is demonstrated through CPU-based simulations and a proof-of-principle run on the Ascella QPU. The results show feasible quantum-assisted classification under current hardware constraints and establish a scalable workflow that can leverage larger quantum resources as photonic QPUs improve. While not claiming quantum advantage today, the study provides a concrete, chemistrydirected path for future quantum-enhanced materials informatics.

Abstract

Polymers exhibit complex architectures and diverse properties that place them at the center of contemporary research in chemistry and materials science. As conventional computational techniques, even multi-scale ones, struggle to capture this complexity, quantum computing offers a promising alternative framework for extracting structure-property relationships. Noisy Intermediate-Scale Quantum (NISQ) devices are commonly used to explore the implementation of algorithms, including quantum neural networks for classification tasks, despite ongoing debate regarding their practical impact. We present a hybrid classical-quantum formalism that couples a classical deep neural network for polymer featurization with a single-photon-based quantum classifier native to photonic quantum computing. This pipeline successfully classifies polymer species by their optical gap, with performance in line between CPU-based noisy simulations and a proof-of-principle run on Quandela's Ascella quantum processor. These findings demonstrate the effectiveness of the proposed computational workflow and indicate that chemistryfrelated classification tasks can already be tackled under the constraints of today's NISQ devices.

Figures (12)

• Figure 1: End-to-end schematic workflow for polymer property prediction using quantum machine learning. Starting from multiscale polymer models (monomer, macromolecular, or mesoscopic scales), monomer structures represented by SMILES strings are converted into numerical vectors and transformed into $n$-dimensional feature vectors through a neural-network-based featurization step, followed by quantum encoding (embedding) as depicted in Fig. \ref{['fig:feature_pipeline']}.The resulting quantum states are then fed into a quantum processing unit (QPU) or a noisy simulator to estimate polymer properties.
• Figure 2: Overview of the polymer‐sequence featurization and encoding pipeline. Monomer sequences are first converted into fixed-length numerical vectors. A classical neural network-consisting of an Embedding layer, a Bidirectional LSTM, a Time-Distributed Dense layer, and a final Dense layer after reshaping-maps these monomer vectors $\bm{x}\in \mathbb{R}^N$ to low-dimensional latent representations $\bm{x}'\in \mathbb{R}^N$. The resulting reduced-dimension monomer embeddings are then passed to the quantum pipeline, where each vector is encoded into corresponding quantum states, enabling hybrid classical-quantum downstream processing.
• Figure 3: Reproduced from Fig. S11 in Ref. Ascella_2023, with modifications: Linear optical QPC implemented on Qunadela’s QPU Ascella for the VQC approach. Spatial modes 3 to 7 are selected out of the 12 available modes to form the first parametrizable (by $\bm{\theta_1}$) block on the QPC by exploiting 16 of the re-configurable phase shifters. The second parametrized (by $\bm{\theta_2}$) trainable block with 16 phase shifters is followed by a pseudo PNR (PPNR) detectors block (in orange) that spreads over four extra spatial modes. By design, four phase shifters acting on spatial modes 4 to 7 are attributed to encode four-feature data block, sandwiched between the two trainable blocks (in grey). The phases of the chip that have neither trainable nor data encoding functions are set to 0.
• Figure 4: Schematic of the learning loop for the photonic variational quantum classifier (modified version of Fig. S11 in Ref. Ascella_2023). A photonic circuit specified by the unitary $U$ is parametrized by ($\bm{\theta}$, $\bm{\lambda}$) fed with the encoded input $\bm{x'}$. The measurement statistics provide the model prediction (output) $f_{\bm{\theta},\bm{\lambda}}(\bm{x'})$, from which the loss $\mathcal{L}(\bm{\theta},\bm{\lambda})$ is computed and used to update the parameters to ($\bm{\tilde{\theta}}$, $\bm{\tilde{\lambda}}$) for the next optimization step. $\bm{\lambda}$ are the variational parameters of the model.
• Figure 5: Confusion matrices for the binary classification of polymers in the dataset comprising 5281 species. Each polymer is represented by a 4D feature vector encoding its monomeric chemical structure. The feature vectors are obtained using the DNN-based feature extractor, either directly with $k$ = 4 or by $k$ = 2 features followed by data augmentation to construct a 4D representation.