Symposium Program PRELIMINARY

Leveraging machine learning for quantum technologies — digital twins and beyond

Full abstract to follow.

Machine Learning and Photonic Quantum Computing

The interaction between machine learning and photonic quantum computing runs in both directions: photonic processors can be used to construct learning models, while classical optimization and machine learning can help design the processors themselves.

I will first introduce MerLin, an open-source framework that brings differentiable photonic quantum circuits into standard PyTorch workflows. An important part of the project is the systematic reproduction of published quantum machine-learning results in a common environment. This allows us to look more carefully at how performance depends on choices such as the dataset, preprocessing, encoding, simulation assumptions and classical baselines. I will share some of the lessons we have learned from this effort and discuss what it takes to move from promising individual demonstrations to evidence that is reproducible, comparable and relevant to real photonic hardware.

In the second part, I will look at the problem from the hardware side: how can computational methods help us design photonic components? Nanophotonic inverse design is a useful example. Much of it relies on physics-based gradient optimization rather than machine learning in the strict sense. However, learned surrogate and generative models can complement these methods, particularly when designs need to be adapted across many target specifications or fabrication constraints.

Closing the Loop on Quantum Dot Control: Real-Time Feedback for Scalable Spin-Qubit Arrays

Semiconductor spin qubits have progressed from few-qubit demonstrations toward increasingly complex quantum-dot arrays fabricated in reproducible academic and industrial processes. As these systems scale, a central challenge is no longer simply finding a good operating point, but maintaining that operating point in the presence of electrostatic drift, charge rearrangements, device variability, and many interdependent controls. I will describe recent work toward feedback-enabled control of semiconductor quantum-dot devices, where machine-learning-based measurement interpretation, image-processing methods, physics-informed heuristics, and automated control routines are combined to detect changes in device state and apply corrective updates in real time.

I will focus on recent results from a 10-dot planar germanium device, including more than 48 hours of autonomous monitoring, detection of electrostatic shifts as small as 0.1 mV, real-time feedback compensation, and spatially resolved noise-correlation measurements. I will also discuss a recent live demonstration of this approach and the broader workflow requirements for deploying feedback across larger arrays. These results complement our modular control stack for quantum-dot devices, where standardized intermediate data products, explicit interfaces between tuning modules, and workflow-level performance metrics enable scalable, autonomous workflows and reliable operation.

AI enhanced infrastructure software: integrating machine learning across the quantum computing stack

Quantum computers promise to revolutionize computing and have made incredible hardware strides, but they remain notoriously error-prone. While low-level control tweaks and clever circuit design can compensate for these hardware errors, current techniques require the availability of highly detailed noise models, deep algorithmic knowledge, and labor-intensive manual tuning. Because of this, they rarely generalize beyond a few narrow use cases. In this talk, I will show how machine learning (ML) can be integrated across the entire quantum computing stack to solve this bottleneck. We will explore how ML methods outperform traditional techniques at every level: from device and gate calibration at the bottom, to compilation and error suppression in the middle, up to algorithmic optimization and quantum error correction at the top.

Can noise help quantum algorithms? Insights from sampling and variational methods

Quantum algorithms typically claim to be resource efficient relative to classical algorithms but it is well known that this performance differential may vanish under realistic operating conditions or noise. While the field has developed strategies to retain quantum advantage, the interaction between quantum and classical resources is often ignored but critically determines the efficacy of quantum algorithms. Our first result, in the context of noisy quantum variational learning, shows that performance depends on how noise interacts with classical optimisation. I show that commonly used subroutines in error‑mitigation strategies such as Pauli twirling can degrade variational optimisation by suppressing gradients and reducing expressivity. In contrast, biased or non‑unital noise can introduce exploitable structure that improves optimisation outcomes. Through analytical and numerical studies, this work demonstrates that preserving noise asymmetries can lead to better performance than symmetrising noise, challenging standard assumptions about mitigation in the variational setting. In recent times, quantum variational models have evolved towards sample-based quantum algorithms for chemistry. For the specific example of Sample‑based Quantum Diagonalisation (SQD), in our second result, I show that noise enables a higher unique sampling rate that allow classical stochastic algorithms to outperform quantum algorithms for naively designed chemistry demonstrations. I discuss that standard error mitigation is not useful for these algorithms, before showing how machine learning can enable better ansatze design and measurement techniques to not only increase the unique sampling rate but ensure noise-robustness and resource efficiency of these demonstrations.

Scalable Neural Decoders for Practical Real-Time Quantum Error Correction

Real-time, scalable, and accurate decoding is a critical component for realizing a fault-tolerant quantum computer. While Transformer-based neural decoders such as AlphaQubit have demonstrated high accuracy, the computational complexity of their core attention mechanism, which scales as O(d⁴) with code distance d, results in decoding speeds insufficient for practical real-time applications. In this work, we introduce and evaluate a Mamba-based decoder, a state-space model with O(d²) complexity. In memory experiments using Sycamore hardware data, our Mamba decoder matches the performance of its Transformer-based counterpart, providing that its superior efficiency does not come at the cost of performance. Crucially, in simulated real-time scenarios that account for decoder-induced noise, the Mamba decoder significantly outperforms the Transformer, exhibiting a higher error threshold of 0.0104 compared to 0.0097. These results demonstrate that Mamba decoders offer a compelling balance between speed and accuracy, making them a promising architecture for scalable, real-time quantum error correction. (arXiv:2510.22724)

Encoded Quantum Signal Processing for Heisenberg-Limited Metrology

Entangled quantum probes can achieve Heisenberg-limited measurement precision, but this advantage is typically destroyed by noise. We address this issue by introducing a framework that we call encoded quantum signal processing, which unifies quantum error detection and quantum signal processing into an effective single-qubit framework, and provides a paradigm for constructing logical sensors that are robust to noise while remaining sensitive to the signal of interest. We show that encoding sensor qubits into a repetition code and using syndrome measurements as a signal-processing primitive restores Heisenberg scaling under realistic noise. The underlying mechanism, which we formalize as encoded quantum signal processing, reduces multi-qubit metrology to an effective single-qubit problem where syndrome measurement implements nonlinear signal transformations. Numerical simulations validate the theoretical predictions: syndrome-based inference achieves near-Heisenberg scaling at noise levels where bare probes approach the SQL, and a concatenated protocol maintains this scaling under joint transverse noise and longitudinal inhomogeneities.

Computational Structure of Learning in Quantum Systems

How can quantum systems learn from data? Quantum learning problems arise naturally when quantum systems are used as models, as data sources, or as physical platforms. This talk examines learning in quantum settings through two complementary perspectives: using quantum systems as learning models, and learning quantum systems themselves from data. On the modelling and designing side, a recurrent theme is the role of structure in determining learnability. Structured operator representations, together with tools from quantum linear algebra, lead to quantum models whose complexity and behavior can be analyzed more systematically. From a complementary perspective, one can view tasks such as reconstructing states, channels, or Hamiltonians as problems of learning quantum systems. This connects naturally to ideas from system identification and operator approximation. I will also discuss recent directions on coherent training protocols, where model parameters are treated as quantum degrees of freedom and optimized through quantum evolution. More broadly, the goal is to better understand what makes quantum systems learnable, and how these principles shape the design of quantum learning and computational models.

Data-Driven Belief Propagation Decoding of quantum LDPC Codes under Realistic Noise

Quantum error correction enables reliable logical computation on noisy hardware, but its guarantees typically rest on simplified noise models and incur large physical overheads. Real devices, however, exhibit correlated errors that fall outside these models. We enhance the performance of quantum low-density parity-check (qLDPC) codes, a family of codes which contain candidates with the lowest known overheads for fault tolerance, under spatially correlated noise. We leverage a correspondence between belief propagation (BP) decoding of a qLDPC code and a neural network, in which the message-passing iterations of BP unroll into the layers of the neural network. This correspondence opens a wide scope for data-driven generalizations of BP. We train the neural network weights on characterization data obtained via Cycle Error Reconstruction, yielding a decoder adapted to the error model of the underlying hardware. Our neural network decoder achieves up to a tenfold reduction in logical error rate over standard BP across diverse noise regimes.

Seeking advantage in Quantum-Classical Machine Learning with trapped ion system

Quantum technologies, specifically Quantum Computing (QC), represent a paradigm shift in computational power. However, the roadmap to utility is hindered by the inherent fragility of quantum states—a hurdle that makes scaling the industry's greatest challenge. To address this, we have developed a quantum-classical hybrid framework designed to push the boundaries of Quantum Machine Learning (QML). By leveraging the trapped-ion platform—noted for its superior coherence times and high gate fidelities—we have demonstrated that quantum classifiers can achieve over 98% fidelity in supervised learning tasks using real-world datasets [1, 2]. While achieving parity with classical systems is a significant milestone, our current research pivots toward the ultimate goal: identifying the specific parameters and architectures where QML provides a definitive advantage over classical counterparts. Here, we will discuss detail of our findings across varied application contexts, highlighting the conditions under which quantum enhancement becomes a reality. [1] 10.1103/physreva.106.012411 [2] 10.1016/j.isci.2025.113058

Machine Learning, Quanta, and the Extraction of Physical Compute Power

Quantum computing's industrial and scientific impact depends on more than established computational advantages; it requires progress across algorithms, architecture, engineering, and foundational concepts. This talk advocates a vertically integrated, interdisciplinary systems perspective. We discuss how machine learning can help extract greater value from quantum systems by reducing unnecessary classical computation in hybrid algorithms and improving use of available quantum resources. We argue that this broader framing can reveal deeper structural origins of quantum advantage and inspire new views of learning.

Artificial Intelligence for Representing and Characterizing Quantum Systems

I review how integrating AI into quantum-system characterisation is typically based on machine learning, deep learning and language models for the core tasks of quantum-property prediction and quantum-system reconstruction with applications to certification, benchmarking, enhancing quantum information processing and identifying critical quantum phenomena.

Quantum advantage of fidelity kernels and molecular descriptors of quantum models

I will first demonstrate that quantum fidelity kernels can be proven to have a quantum advantage, in principle. How to take advantage of this quantum advantage in practice remains unclear. To address this challenge, I will describe several algorithms for building high-performing quantum circuits for specific machine learning tasks. I will discuss the isomorphism between quantum circuits and a subspace of polyatomic molecules and argue that molecules can be used as descriptors of quantum models. Finally, I will present an algorithm for improving quantum models using the Bayesian information criterion as well as the dimension of the dynamical Lie algebra as model selection metrics.

Learning Quantum State Preparation with Generative AI

Preparing quantum states efficiently remains a central challenge in quantum chemistry. This talk presents recent work on using generative AI to learn molecular ground-state preparation circuits directly from molecular Hamiltonians, replacing expensive iterative circuit construction with learned generation. I will discuss the insights gained from this approach, its current capabilities, and limitations, and conclude with an outlook on generative quantum eigensolvers and future directions for machine learning in quantum algorithm design.

Beyond Optimization: Quantum Annealing for Generative Machine Learning

Quantum annealers are commonly viewed as optimization engines, but their ability to sample from complex energy-based distributions provides another natural interface with machine learning. In this talk, I will describe a hybrid architecture in which a quantum processing unit trains and samples an energy-based prior embedded within the discrete latent space of a classical generative model.

Using molecular generation as a case study, I will discuss how continuous neural representations are mapped to binary latent variables, how a graph-restricted Boltzmann-machine prior is trained using quantum annealing, and how samples from the learned distribution are returned to a classical decoder. Compared with the classical sampling baselines studied, the QPU-trained models generated molecules with higher validity and drug-likeness.

Rather than replacing the classical ML model, the quantum processor addresses a specific sampling task within it. I will conclude with lessons from this work on where quantum sampling may fit within existing ML workflows, how these hybrid approaches should be benchmarked, and what makes a problem a promising candidate for further exploration.

Toward ML for Quantum Algorithms at Practical Scale

How to apply modern machine learning techniques to quantum algorithms at practically relevant scales remains an open question. We will explore this question through the Generative Quantum Eigensolver (GQE) and its variants, which use generative machine learning to discover quantum circuits for quantum chemistry. We will then discuss MoLe, an approach for constructing molecular wave functions that can transfer across different molecules, highlighting a path toward scalable and transferable machine learning for practical quantum algorithms.

Data-enhanced self-learning Projection Quantum Monte Carlo

Projection quantum Monte Carlo (PQMC) provides unbiased ground-state estimates for stoquastic quantum many-body Hamiltonians, but its practical efficiency depends strongly on the quality of the guiding wavefunction used for importance sampling. We introduce a data-enhanced self-learning PQMC framework in which projective measurements from a programmable Rydberg-atom quantum simulator are used to pretrain neural guiding wavefunctions before the standard self-learning refinement. We apply the method to the square-lattice Rydberg Hamiltonian across the disordered-to-checkerboard transition, using both restricted Boltzmann machine (RBM) and autoregressive recurrent neural-network (RNN) ansätze. Pretraining on experimental measurement data substantially improves the initial guiding distribution, accelerates convergence of the mixed energy estimator, reduces finite-walker bias, and lowers the number of walker and training iterations required to reach accurate ground-state energies. The improvement is especially pronounced for RBM-guided wavefunctions, where data pretraining mitigates convergence to poor local minima, while RNN-guided wavefunctions benefit from faster convergence and improved sample efficiency. We further show that the advantage persists across system sizes, with the residual energy exhibiting a weaker walker-number scaling than in conventional self-learning PQMC. These results demonstrate that imperfect but physically informative quantum-simulator data can be used as a resource to enhance classical projector Quantum Monte Carlo simulations, suggesting a practical hybrid route for accurately studying quantum matter with near-term quantum devices.

Grokking and epoch-wise double descent in quantum neural networks

Grokking, the delayed transition from memorization to generalization, is a fundamental phenomenon in gradient-based learning, yet its dynamics within variational quantum machine learning (QML) remain largely unexamined. In this work, we report the empirical observation of both the grokking transition and epoch-wise double descent in a two-qubit quantum neural network (QNN) under a complete parameterization of the SU(4) manifold. We demonstrate that overparameterization via increased circuit depth improves the probability of successful generalization. Notably, these architectures frequently exhibit an epoch-wise double descent in test error, degrading at a critical epoch before recovering into a generalizing state. Crucially, we identify a generalization decay in late-stage training, where the test error increases significantly despite a stagnant training loss. Bridging this behavior with algorithmic stability theory, our analysis reveals that this decay correlates with an unconstrained increase of the weight-norm, drifting away from sparse, phase-aligned harmonic solutions toward overfitted solutions in the Hilbert space. We analyze the underlying temporal dynamics of this transition, demonstrating how the onset of generalization is linked to optimization hyperparameters such as learning rate and weight decay. Finally, to mitigate late-stage decay, we introduce a weak explicit weight-norm regularization into the loss function. We demonstrate that this structural anchor stabilizes the post-grokking phase and permanently preserves generalization gains, providing a robust framework for training overparameterized quantum circuits.

Machine Learning for Autonomous Characterization and Control of Semiconductor Quantum Devices

As semiconductor quantum processors continue to increase in complexity, the effort required to characterize, calibrate, and operate these devices is becoming a fundamental bottleneck to scalability. While advances in fabrication have enabled larger quantum dot arrays, many experimental workflows still rely on manual interpretation of measurement data and expert-guided tuning. Machine learning offers a promising path toward autonomous quantum hardware by enabling rapid interpretation of experimental signals, adaptive decision making, and closed-loop optimization.

In this talk, I present a machine learning framework for autonomous characterization and control of semiconductor quantum dot devices, spanning spin-qubit initialization and electron-pump operation. The approach combines physics-based simulation with supervised deep learning to infer device state directly from transport measurements, including charge stability diagrams and charge pumping maps. Convolutional neural networks and U-Net architectures are trained on large synthetic datasets generated from electrostatic and transport models, enabling robust identification of charge transition lines, quantum dot configurations, and quantized pumping plateaus across a wide range of realistic noise conditions.

These learned representations form the foundation of autonomous tuning algorithms that replace manual feature extraction with rapid, reliable inference. By integrating machine learning with physical models and experimental feedback, this work demonstrates a scalable strategy for automated quantum device calibration while providing a pathway toward adaptive control of larger semiconductor quantum processors. More broadly, the framework illustrates how combining domain-specific simulations with modern computer vision techniques can accelerate the development and operation of next-generation quantum technologies.

Ravines in quantum cost landscapes: opportunities for improved VQA predictions

The geometric and topological structure of quantum cost landscapes (QCLs) governs the optimization and thus the predictive power of variational quantum algorithms (VQAs). We systematically analyze ravines — low-cost paths connecting local minima — using an adapted version of the nudged elastic band (NEB) algorithm, a method originating from theoretical chemistry. By training quantum neural networks (QNNs) to classify the concentratable entanglement of quantum states, we apply the NEB algorithm and numerically identify ravine structures in QCLs of hardware-efficient ansatzes. Beyond visualizing these ravines, we construct an ensemble prediction framework by averaging predictions from QNNs parameterized along the low-cost NEB path. We introduce a resource-light pre-training metric which quantifies local prediction variability and serves as a strong performance indicator for VQAs, even beyond the scope of this study. When base classifiers are drawn from circuit and weight initializations exhibiting high local-prediction variability, the quantum-based NEB ensembles outperform both classical and naive quantum alternatives. Moreover, a complexity analysis shows that leveraging the ravine-like structure of QCLs with the QNN NEB approach substantially reduces computational costs compared to naive QNN ensembling. A depth and qubit scaling analysis indicates that ravines persist across both scalings, and that, despite the expected growth in resource requirements with the qubit scaling, the NEB approach also accelerates convergence over the naive alternative. (arXiv:2607.01329)

Learning fermionic linear optics with Heisenberg scaling and physical operations

Fermionic linear optics (FLO), equivalent to fermionic Gaussian unitaries or matchgates on a line, have wide applications ranging from device benchmarking, classical shadows, quantum neural networks, and more. We revisit the problem of learning FLO circuits: given black-box query access to an unknown N-qubit FLO, produce an approximate description using as few resources as possible. Previous proposals featured 1) N^5 query complexity; 2) standard quantum limit scaling in precision; 3) unphysical operations violating fermionic superselection rules; and 4) N auxiliary quantum space. In this work, we establish efficient and experimentally friendly protocols that address all of these deficiencies: 1) we improve to N^4 scaling in general, which is further reduced to N^3 for number-conserving (passive) FLOs; 2) we achieve the optimal Heisenberg limit; 3) all operations obey superselection; and 4) we only require 1 ancilla qubit, which can be removed in most physically-relevant scenarios. This brings the task of learning FLOs closer to practical implementation, and marks the first such algorithm that attains Heisenberg scaling.

This work is joint with Aria Christensen. A preprint is available at: https://arxiv.org/abs/2602.05058

Exact Stochastic Schrödinger Equations for Quantum Reverse Diffusion

A few years ago, the field of machine learning (ML) and, more broadly, our culture were shaken by advances in diffusion-based generative models such as Midjourney, DALL-E, Sora, Veo, etc. The breakthrough in the generative capabilities can be attributed mainly to the 40-year-old mathematical theory of classical reverse diffusion by B. D. Anderson. He showed that a Markov diffusion process can have a reverse diffusion process, e.g., noisy data dynamically and stochastically evolve into noise-free data.

Following advances in classical ML, the field of quantum generative modelling began to gain traction and expand. The core idea of quantum generative modelling is to define a quantum analog of a naturally occurring noisy forward process and then use hybrid techniques, including variational quantum circuits, to imitate the quantum reverse process. While these quantum ML techniques are powerful for learning a surrogate of the reverse, they do not reveal which physical principles fundamentally define a natural quantum reverse process, which, by definition, must incorporate the same noise and decoherence effects as the forward process. In other words, most current works do not realize the reverse process and instead imitate it using variationally learned and fully coherent dynamics. Although these numerical schemes imitate quantum reverse dynamics, analytical equations describing a physical quantum reverse diffusion process, on the same footing as the forward stochastic Schrödinger equation (SSE), have so far been absent.

In our work, we provide a rigorous theoretical foundation for quantum reverse diffusion in continuously monitored quantum systems. Specifically, for forward quantum stochastic processes driven by monitored Pauli noise, we derive a family of exact stochastic Schrödinger equations that describe the corresponding reverse quantum processes. These reverse stochastic Schrödinger equations are generalizations of the forward SSE and, as such, preserve the noise and decoherence structure of the forward dynamics. Furthermore, the reverse processes are mathematically guaranteed to reverse the noise effects almost surely. The stronger almost sure reversal of the forward dynamics can be relaxed by configuring the reverse process to steer the state onto a manifold of states; in this case, the dynamics implement a reversal in the distribution. Therefore, the presented reverse SSEs are powerful, as they enable a wide spectrum of applications, from almost sure state recovery to quantum generative modelling.

Local tensor-train surrogates for quantum learning models

A key bottleneck in quantum machine learning is the computational cost of repeated quantum circuit evaluations during the inference phase. To address this, we present a framework for constructing fast, cheap, provably accurate classical tensor-train surrogates of fully trained quantum machine learning models within local patches of their input data space. The approach combines Taylor polynomial approximation with a tensor-train (TT) representation and embeds it in a statistical learning paradigm via empirical risk minimization. In our analysis, the Taylor-TT construction serves as a deterministic error certificate proving that the TT hypothesis class contains a good approximation; empirical risk minimization then provably recovers a surrogate with controlled generalization error and explicit bounds. This translates into three independently controllable error sources: (i) Taylor truncation error controlled by the patch radius $r$ and polynomial degree $p$, (ii) TT approximation error controlled by the bond dimension $\chi$, and (iii) statistical estimation error. While the parameter count scales polynomially in the number of data dimensions $N$, i.e., $\deff = N(p+1)\chi^2$ rather than the naive $(p+1)^N$, the worst-case constants inherit an exponential factor through the tensor-product feature norm during Taylor polynomial embedding onto TT. This cleanly separates representation complexity from feature-induced constants. Our risk bounds and sample complexity depend explicitly on the local patch radius $r$. the arXiv preprint is available at https://arxiv.org/abs/2604.25631.

Generative Learning of Optimal Quantum Measurements

Efficiently measuring quantum states is a central challenge in quantum computing. In many applications, we need to estimate a large number of Pauli observables, many of which do not commute, while using as few measurements as possible. A standard approach is to group observables into commuting sets and measure each group together. However, general commuting groups can require relatively deep basis-change circuits, which are difficult to implement on near-term devices and may still be costly in early fault-tolerant settings. A more hardware-friendly option is to use only single-qubit rotations and qubit-wise commuting groups. This keeps the measurement circuits shallow, but usually produces smaller groups and therefore increases the number of samples needed. In this talk, I will present a new measurement scheme based on generative learning. Instead of explicitly constructing measurement groupings through NP-hard combinatorial optimization, or relying on complex derandomization procedures for classical shadows, our method directly learns an ensemble of measurement circuits tailored to a target set of observables and practical resource constraints, such as the total measurement budget and allowed circuit depth. Our numerical results show systematic improvements over state-of-the-art measurement strategies, as well as encouraging generalization beyond the training regime. Overall, this learning-based approach offers a flexible framework for practical measurement design in both near-term and early fault-tolerant quantum computing, where quantum resources remain limited and imperfect.

Trainability and Mode Seperation of Mixed IQP circuits

Instantaneous quantum polynomial-time (IQP) circuits are a promising route to generative modeling in the noisy intermediate-scale quantum (NISQ) era, supporting a train-on-classical, deploy-on-quantum paradigm: the low-body Pauli-Z expectation values that assemble a Maximum Mean Discrepancy (MMD) loss are classically estimable, while sampling from the trained circuit remains classically hard in the relevant regime -- so training is fully classical, yet generation retains a genuine quantum-advantage target.

This promise is limited by a tension between expressivity and trainability. An ancilla-free IQP circuit is not universal -- even on two qubits there are distributions no choice of angles reproduces -- yet restoring universality by appending an ancilla register reinstates the barren plateaus that obstruct training. We resolve this tension by working in the decomposed picture of an ancilla-augmented IQP circuit: a weighted mixture of ancilla-free IQP circuits (branches) that share one interaction graph but carry independent angles, which we call a mixed IQP.

Using the mixed IQP, we identify three data-informed initializations that keep the ancilla-augmented circuit trainable, and prove that its training succeeds only when the branches are seeded to carry distinct distributions. We show that the mixture inherits the barren-plateau avoidance of a single circuit for a polynomial number of branches, with only a polynomially small suppression of the loss curvature. Building on this, we introduce a data-partitioning cluster initialization that seeds each branch from a distinct data mode, prove it is trainable, and show it supplies the inter-branch diversity a mixture needs to surpass the ancilla-free performance -- whereas branches that collapse to near-identical distributions provably cannot, and a coincident start receives no first-order gradient to differentiate them. Branch weights are promoted to a free control set through the ancilla state, and the trained model is deployed by randomized per-shot selection of a single circuit, with no ancilla overhead at sampling time.

We demonstrate experimentally that adding ancilla branches to the mixed IQP raises its performance well beyond that of an ancilla-free IQP circuit, across four benchmarks spanning system sizes from sixteen to seven hundred eighty-four qubits and three interaction-graph families. This gain appears only when the data separate into distinct modes: on well-separated targets, routing each branch to its own mode lifts performance, while on the largest image target, whose modes overlap in Hamming space, closely spaced modes need only a small inter-branch separation to be told apart -- one the model installs on its own from any initialization, so every scheme fits the data equally well, and the mixture's gain tracks the mode structure of the data rather than the system size. Through the mixed IQP, we thus establish -- both theoretically and experimentally -- the successful training of ancilla-augmented IQP circuits.

AI and the Future of Scientific Research

As large language models and AI research agents become increasingly capable, how should scientific research and the role of the scientist change? This panel explores human contributions to AI-enabled discovery, including choosing meaningful questions, curating results, and navigating tensions between scientific progress and career incentives. We will discuss how researchers and journals can verify findings, assess significance, and maintain accountability as output grows. Finally, we will consider how graduate training and supervision should evolve to develop independent scientific judgment. Grounded in theoretical physics and quantum research, the discussion will include audience questions and forecasts for the years ahead.

Geometry-Induced Long-Range Correlations in Recurrent Neural Network Quantum States

Neural Quantum States based on autoregressive recurrent neural network (RNN) wave functions enable efficient sampling without Markov-chain autocorrelation, but standard RNN architectures are biased toward finite-length correlations and can fail on states with long-range dependencies. A common response is to adopt transformer-style self-attention, but this typically comes with substantially higher computational and memory overhead. Here we introduce dilated RNN wave functions, where recurrent units access distant sites through dilated connections, injecting an explicit long-range inductive bias while retaining a favorable \mathcal{O}(N \log N) forward pass scaling. We show analytically that dilation changes the correlation geometry and can induce power-law correlation scaling in a simplified linearized and perturbative setting. Numerically, for the critical 1D transverse-field Ising model, dilated RNNs reproduce the expected power-law connected two-point correlations in contrast to the exponential decay typical of conventional RNN ansätze. We further show that the dilated RNN accurately approximates the one-dimensional Cluster state, a paradigmatic example with long-range conditional correlations that has previously been reported to be challenging for RNN-based wave functions. These results highlight dilation as a simple geometric mechanism for building correlation-aware autoregressive neural quantum states.

Random Telegraph Signal Analysis for Semiconductor Quantum Device Characterization: Machine-Learning-Assisted and Training-Free Approaches

Random telegraph signals (RTSs) provide a direct time-domain probe of stochastic switching processes in semiconductor devices. In quantum-dot, single-photon detector, transistor, carbon nanotube, and related platforms, RTS features can reflect charge trapping, tunneling, defect activity, and other microscopic processes that affect device performance. However, robust RTS analysis remains difficult when measurements contain strong background noise, multi-level switching, overlapping traps, or large parameter sweeps. Manual inspection can also make the analysis hard to reproduce.

This poster presents an automated RTS analysis pipeline for semiconductor quantum device characterization. The pipeline is designed as a modular workflow that can combine denoising, digitization, dwell-time extraction, and statistical interpretation. Within this framework, Denoising Autoencoder based on U-Net and bidirectional LSTM is used as one promising machine-learning denoising method for complex RTS traces. Adaptive dual-tree complex wavelet transform denoising followed by Bayesian digitization is used as another promising combination, especially when training-free operation, speed, and reproducibility are important. The goal is not to promote a single universal method, but to build a practical analysis toolkit in which different methods can be selected based on the measurement regime and the downstream physics question.

We apply this RTS workflow to semiconductor quantum-dot photon-arrival data. Conventional dense-sample analysis reproduces millisecond-scale charge tunneling dynamics reported in earlier optical charge-state measurements. Targeted dense-sample analysis then focuses on bright-state intervals and resolves faster microsecond-scale structure that is hidden in an effective two-state charge model. In parallel, single-photon detection interval analysis uses the raw detector timing resolution to examine photon-spacing distributions without first reducing the data to an RTS trace. This gives access to nanosecond-to-microsecond optical and emission dynamics.

Together, these results show how ML-assisted and automated signal analysis can support quantum hardware characterization. The work connects noisy time traces to physical time scales across semiconductor platforms, and it points toward reproducible high-throughput workflows for device diagnostics, parameter scans, and quantum-device reliability studies.

Note: This poster is part of a pair of closely related submissions from our group, named Hardware-Aware Quantum Machine Learning Pipelines for Inference with Quantum Processing Unit by Bowen Deng. Two of us (Tonghe Bai, Bowen Deng) are deeply involved in both projects, and we have assigned one primary contact to each poster for the submission form. If possible, we would appreciate having the two posters placed beside each other, so both presenters can participate in discussions for both posters.

High-dimensional entanglement witness certificate with machine learning methods

The Positive under Partial Transposition criterion represents the first level of the hierarchy in a complete family of separability criteria based on the PPT symmetric extension of the quantum state $\rho$. It holds for bipartite systems whose total dimension $d \leq 6$. For $d > 6$, there's no analytical criterium able to detect entanglement for any bipartite or multipartite state taken into consideration. The entanglement detection problem can be formulated as a numerical optimisation problem. This motivates the exploration of numerical methods - in particular, machine learning methods - to develop computational solutions towards this goal. While most of the approaches in the machine learning literature detect entanglement up to a certain accuracy, we develop a deep learning method that given a quantum state $\rho_e$ constructs a certifiable entanglement witness $W$ using the state's PPT symmetric extension above mentioned. Our method shows a higher efficiency in scaling the dimension of the system with respect to the computational resources used by the analogous Semi-Definite Programming formulation.

Hardware-Aware Quantum Machine Learning Pipelines for Inference with Quantum Processing Unit

Quantum machine learning (QML) models are often developed in ideal simulators, but practical use on quantum processing units requires a second layer of design. A model must be trainable in simulation, mappable to available hardware, executable under realistic noise and connectivity constraints, and interpretable enough to diagnose why performance changes after deployment. This poster presents a hardware-aware QML workflow built around these requirements.

The first component is simulator-based QML training and evaluation. We study hybrid classical-quantum models for electrocardiogram beat classification using a staged architecture: classical feature extraction, quantum feature encoding, and quantum or classical classification. This structure makes it possible to compare classical and quantum baselines under controlled conditions. The strongest current configuration uses a classical feature extractor, amplitude-based quantum encoding, and an angle-based quantum classifier, reaching mean 95% test accuracy in the noiseless simulator setting.

The second component is the mapping of trained QML circuits toward real Quantum Processing Unit (QPU) inference. Rather than treating hardware execution as an afterthought, the workflow tracks whether the trained circuit can be expressed using hardware-compatible gates, whether the entangling pattern matches device connectivity, and how circuit structure changes under transpilation. This includes tests under realistic noise models such as live IBM backend QPUs.

The third component is qubit-patch selection and routing. For a fixed quantum model size, multiple connected qubit subsets may be available on a QPU. These patches differ in two-qubit error rates, readout errors, coherence times, and routing overhead. We therefore rank candidate qubit patches, test whether the circuit stays inside the selected patch, and check whether swaps or unwanted routing are introduced. This gives a more concrete path from trained QML model to hardware inference batch.

The fourth component is topology-efficient trainable entangling circuit design. Instead of relying only on generic all-to-all or simulator-friendly ansatz, we study trainable entangling circuits whose two-qubit gates follow the native hardware graph. This reduces unnecessary routing and makes the learned circuit more compatible with real QPU constraints.

Overall, this work fits the ML4QT software theme by connecting QML algorithm design, circuit synthesis, hardware-aware inference, and topology-constrained deployment. The central question is not only whether a QML model can classify data in simulation, but how much of that model remains usable when it is prepared for real quantum hardware.

Note: This poster is part of a pair of closely related submissions from our group, named Random Telegraph Signal Analysis for Semiconductor Quantum Device Characterization: Machine-Learning-Assisted and Training-Free Approaches by Tonghe Bai. Two of us (Bowen Deng, Tonghe Bai) are deeply involved in both projects, and we have assigned one primary contact to each poster for the submission form. If possible, we would appreciate having the two posters placed beside each other, so both presenters can participate in discussions for both posters.

PadoPauli: A GPU-Accelerated Implementation of Pauli Propagation

Pauli propagation casts evolving observables as propagating strings of Pauli operators instead of evolving quantum states. Given a parameterized circuit with a fixed gate architecture, this formulation reduces evolving observables to a fixed architecture of Pauli strings with parameterized coefficients, with no need for expensive state computations. Existing CPU and GPU implementations of Pauli propagation possess limitations such as slow speed and memory growth. These limits can be addressed with a single, reusable compiled propagation structure rather than handling each new parameter as a separate evaluation.

We present PadoPauli, a GPU-accelerated Pauli propagation software for repeated expectation value estimation and optimization. PadoPauli compiles parameterized circuits as Pauli strings in reusable sparse tensor maps. The compiled structure is adjustable with a truncation threshold for computational feasibility. Then, PadoPauli applies what we call zero-filtering to remove Pauli strings that contribute exactly zero to the final expectation value. Zero-filtering reduces the working set by orders of magnitude without accuracy loss. As a tool for quantum machine learning, PadoPauli further adds a differentiation function to the compiled structure.

We validate the applicability of PadoPauli with serveral benchmarks of interest. As an example, compared to cuPauliProp, PadoPauli achieves up to 9.7× faster expectation value evaluation and up to 14.5× faster forward and backward training, while using 21–50× lower peak memory. For batched data embeddings, PadoPauli sustains ≈114× higher throughput at 9 qubits. The compiled structure also closely replicates the 127-qubit kicked Ising IBM Eagle result. As a demonstration of utility, we apply PadoPauli to the training of ma-QAOA. By implementing max-weight truncated Pauli propagation with PadoPauli, we reduce QPU workload by up to 94.5% compared to standard ma-QAOA. These results show that PadoPauli is practical, high-throughput and scalable, suitable as a workflow layer for variational quantum algorithms.

The public version of PadoPauli can be found at https://github.com/Norma-Q/Pauli-Propagation---GPU-acceleration.

Design and Construction of QLDPC Codes on a Two-Dimensional Tweezer Array Using Reinforcement Learning

Fault-tolerant quantum computation requires robust protection against physical noise. Quantum error correction (QEC) provides this by encoding logical qubits redundantly into many physical qubits. Surface and color codes are attractive QEC schemes due to their high error thresholds and locally measurable stabilizers. However, both families encode only a constant number of logical qubits regardless of code size, leading to vanishing encoding rates and poor large-scale scalability. Quantum low-density parity-check (QLDPC) codes overcome this limitation by achieving high, constant encoding rates while retaining sparse, low-weight stabilizers. Their primary challenge is the non-local connectivity their check structure demands.

Neutral atom quantum computers are well suited to this challenge, offering reconfigurable qubit arrangements, long coherence times, and tunable long-range interactions via Rydberg excitations. The platform considered here is a 20×100 two-dimensional tweezer array with independently adjustable columns, where each column of 50 traps can be displaced horizontally, providing a hardware-native mechanism to relax connectivity constraints.

We present a QLDPC code construction framework tailored to this adjustable-column architecture. By designing error correction rounds as alternating sequences of stabilizer measurements and controlled column shifts, we derive a block-structured parity-check ansatz and corresponding CSS orthogonality constraints, and we demonstrate representative code families, including surface-like, concatenated, orthogonal, and self-orthogonal constructions, whose non-local check structure can be mapped directly onto the platform's native operations. We introduce a code design tool that jointly enforces hardware connectivity constraints and the mathematical requirements of the target codes. For a systematic exploration of the code space we trained a reinforcement learning agent to perform a reward-driven search, prioritizing hardware-compatible connectivity patterns and favorable code properties.

Quantum-derived molecular fingerprints using Hamiltonian simulation for data-driven coupled-cluster approach

In quantum chemistry, Coupled-Cluster theory is used to solve the electronic Schrödinger's equation and obtain accurate energies of molecular systems. Solving the coupled-cluster equations requires iteratively finding the optimal one-electron (t1) and two-electron (t2) excitation amplitudes. However, obtaining these parameters scales greater than O(N6), creating a significant computational bottleneck that limits the method's applicability to large molecular systems. To address this, the Data-Driven Coupled-Cluster (DDCC) approach was developed to accelerate the computation of t1 and t2 amplitudes by utilizing a machine learning framework that only requires input features derived from relatively cheaper quantum chemistry methods to directly predict the optimal excitation amplitude values. These predicted excitation amplitudes can then be used to efficiently obtain accurate correlation energies at the Coupled-Cluster Singles and Doubles (CCSD) level of theory by either serving as an optimized initial guess for the iterative process (exact CCSD) or be introduced directly into the energy expression for CCSD (approximate CCSD). In this context, Quantum Machine Learning (QML) models offer a powerful alternative to classical models in the DDCC approach for predicting these amplitudes and aligning with the quantum-centric supercomputing paradigm. Since the performance of any ML model is heavily dependent on its input features, we are developing Hamiltonian simulation based algorithms as a means to generate high-fidelity quantum-derived input features, or quantum fingerprints. Our work explores the idea of feeding these quantum-derived features into QML models to predict the t1 and t2 amplitudes and benchmark the performance for both exact and approximate CCSD correlation energies. Our approach has the potential to enable the estimation of CCSD-quality energies with significantly reduced iterative overhead and providing a methodology for high-level molecular property predictions on emerging quantum architectures.

Error learning with Neural Network decoding dramatically reduces logical error rate

One of the central challenges in building a utility-scale fault-tolerant quantum computer is the large resource overhead required to achieve low logical error rates. Reducing this overhead not only depends on designing improved quantum codes, but also on decoders that can exploit realistic information about hardware noise. Standard belief propagation decoders for quantum low density parity check codes typically assumes local and independent error models, neglecting spatially correlated errors that can arise due to crosstalk, shared control lines and residual interactions which are extremely common in many experimental platforms. Our work overcomes this limitation via a neural-network based variant of belief propagation that incorporates correlated information obtained from cycle-error reconstruction (CER). Our results demonstrate an order of magnitude suppression of logical error rate compared to the state-of-the-art belief propagation-ordered statistics decoding (BP-OSD). We observe this improvement across two variants of qldpc codes such as the hypergraph product (HGP) code and the bivariate bicycle (BB) code. These results suggest that experimentally characterized error correlations can be directly leveraged to substantially improve decoding performance without incurring significant resource overhead and hence provides a promising route towards hardware aware decoding strategies for fault tolerant quantum computing.

Foundations of Whatifiness: Asking Counterfactual What if? Questions through Quantum Principles for Adaptive Decision-making

This research serves as a proof of concept for a quantum reinforcement learning algorithm that can ask What if? questions to adapt to unfamiliar scenarios. To evaluate this property, we train all our control and experimental models on a predefined environmental configuration based on the Minigrid package. This training process is followed by a testing stage in which each model is deployed across a collection of Minigrid environment configurations that differ from those used during training. The adaptability of the models evaluated in our experiment is measured by the reward scores they receive across several episodes on each test grid. In this investigation, we introduce a novel classical-quantum hybrid reinforcement learning method that uses simulated decoherence to ask What if? questions by creating representations of alternative scenarios. Our technique employs an on-policy actor-critic model in which the actor network's quantum layer is based on a tree tensor network circuit, while the critic network is a classical-quantum hybrid transformer model with quantum self-attention layers. Counterfactual samples are generated through random bit flip operations in the quantum layers of the critic network. Each sampling produces a corresponding value function estimate from the critic network; these estimates, along with the value function estimate for the base inputs, are then used to compute a refined estimate for optimizing the actor network. Our transformer critic architecture differs from more traditional critic network structures by using the most recent policy representation from the actor network, in addition to environmental observation data. This feature supports the three counterfactual exploration approaches evaluated in this experiment: the first approach generates alternative policy representations; the second generates modified versions of environmental observation data; the third integrates counterfactual sampling of both observation and policy data. The control component of our experiment consists of two stages: the first stage is characterized by the use of a standard deep neural network architecture consisting of both classical and quantum layers; in the second stage, the critic network from the first stage is replaced by the same transformer architecture as that of the experimental models, but without the use of counterfactual sampling. The model implemented with observation-focused counterfactual sampling achieved the highest performance, demonstrating a significant advantage in adaptability over all other models evaluated in our experiment. This finding supports the viability of simulated decoherence as a mechanism for asking What if? questions in a manner that allows intelligent agents to construct representations of alternative versions of the world. The second principal observation of our experiment is that in the observation-focused counterfactual simulation approach, the use of entangling operations was crucial to the agent's ability to adapt successfully to unseen environmental configurations. This result suggests that the distinctly quantum properties of our counterfactual simulation approach were central to its success at providing the agent with improved adaptive capabilities. The findings of our research gave rise to the concept of Whatifiness as a measure of an intelligent agent's ability to perform causal inference at the counterfactual level, enabling successful problem-solving in unfamiliar scenarios.

Developing quantum centric workflows for accelerating quantum chemical simulations

In this poster, we present the development of a quantum-classical pipeline which employs quantum machine learning (QML) models and a hybrid quantum algorithm to accelerate molecular electronic structure simulations. This work investigates a quantum–centric supercomputing (QCSC) paradigm for accurate predictions of quantum mechanical properties, where quantum computing resources are integrated with traditional high-performance computing workflows to enhance computational simulations. This work incorporates the data-driven coupled-cluster (DDCC) approach which uses the Hartree–Fock derived properties of the wavefunction as input features to predict the t2 excitation amplitudes of the coupled-cluster equations. These amplitudes are computationally expensive to obtain using iterative solvers. We present our investigations on the development of QML models, particularly quantum kernel learning to accurately predict the t2 amplitudes. We also discuss the exploration of other QML models such as quantum neural networks. We further investigate how the QML predicted t2 amplitudes can be used to efficiently calculate the ground state energies of molecules using iterative coupled–cluster solvers. Keeping in view of the QCSC paradigm, we explore the extension of this workflow by utilizing the predicted t2 amplitudes to determine the ground state energies of molecules with a hybrid quantum-classical iterative solver algorithm known as the Sample-based Quantum Diagonalization (SQD) algorithm. We discuss the development and implementation of this solver on real quantum devices using quantum state preparation methods such as the Local Unitary Cluster Jastrow (LUCJ) family of ansatzes.

Experimental adjudication between causal frameworks through machine learning methodologies

Bell's inequalities [1] have proven to be a simple and effective method to test whether nature can be described by classical physics. Experiments which violate these inequalities reveal that the world is correlated in a way that classical physics and classical probability cannot represent. These correlations are referred to as quantum correlations. Bell inequality experiments [2-5] offer a scheme to detect quantum correlations but rely on two key conditions, locality and realism. Locality enforces that experiments performed separately cannot instantaneously influence each other, and realism enforces that physical systems have definitive properties independent of observation. Bell inequality violations show that observed correlations cannot be explained by any classical theory that upholds these conditions. There exists a common misconception that a Bell inequality violation implies that quantum theory must be correct, rather it only implies that at least one of locality and realism must be relaxed for a classical theory to reproduce some observed correlation. To adjudicate between these theories, each can be represented as a causal framework, with the corresponding probability distribution transcoded into a machine learning model, which is trained and then tested on data obtained from a prepare-and-measure experiment. This methodology allows for the detection of overfitting [6]; if a model trains well but tests poorly, it is akin to using a high degree polynomial to recreate data generated from a linear function. This research demonstrates that the additional complexity granted to an exotic theory does not allow it to truly represent the quantum correlations of the system but allows it to fit to experimental noise.

Spectral Recovery Learning for Measurement-Free Quantum Error Correction

We introduce Spectral Recovery Learning (SRL), a spectral framework for measurement-free quantum error correction (QEC) based on quantum autoencoders (QAEs). Instead of viewing QEC solely as a state recovery problem, SRL interprets recovery through the Pauli Transfer Matrix (PTM) representation of quantum channels, where noise redistributes and attenuates information across Pauli modes. From this perspective, a learned recovery channel acts as a spectral reconstruction operator, restoring the physically recoverable components of the noise-free Pauli spectrum by filtering and mixing noisy Pauli modes. This insight motivates both a globally entangling QAE architecture for efficient Pauli-mode mixing and a spectral loss function based on code space restoration and logical information preservation as an alternative to global fidelity optimization.

SRL yields markedly better training behavior than fidelity-based objectives: its predominantly local structure produces a sub-exponential gradient-variance decay and curved, trainable loss landscapes, in contrast to the exponential barren plateau of the fidelity loss, while requiring only a small, fixed set of measurement settings (three circuits for CSS codes). Numerical experiments on repetition, Steane, perfect, and LNCY codes show competitive or superior recovery performance across diverse noise models. For Pauli noise, including bit-flip and depolarizing channels, SRL matches conventional stabilizer-based QEC. Under correlated Gaussian dephasing with nonzero-mean coherent rotations, SRL exploits its single-qubit rotation layers to undo the coherent over-rotation, surpassing syndrome-based recovery, approaching the Petz map at weak noise and exceeding it at stronger noise. For amplitude damping noise, SRL achieves recovery performance comparable to the specialized LNCY code. Furthermore, a recovery model trained at a single noise strength generalizes across all error rates without retraining.

These results establish spectral recovery learning as a physically motivated framework for variational QEC, demonstrating how spectral insights into quantum channels can guide the co-design of recovery architectures and training objectives.

Informed Kernel Quantum Support Vector Machines

Conventional Quantum Support Vector Machines rely on HHL and techniques originally developed for quantum principal component analysis to train least squared-support vector machines. As a result, implementing even a simple QSVM is prohibitively expensive, requiring quantum phase estimation on an operation which queries the full dataset. We propose a QSVM method which operates on classical data with multivariable polynomial kernels. Our method requires an oracle which categorizes each data-point as either +1 or -1, and a similarity function which describes the similarity of two data-points as a polynomial of their similarities on individual variables. Unlike the traditional implementation of HHL, we leverage the structure of our kernel to efficiently calculate the eigenvalues of the training matrix. As a result, we eliminate the much of the complexity of HHL, excluding the projection step. In the long term, we hope to leverage our QSVM for quantum reinforcement learning which will avoid the need for quantum RAM.

Quantum-QKD-Aero: An Open Research Platform for Adaptive Machine Learning in Quantum Communication

Advancing machine learning for quantum communication requires more than optimizing individual protocols—it requires research environments capable of supporting increasingly adaptive communication systems. Quantum-QKD-Aero is a modular, first-principles research platform designed to investigate communication across fibre, free-space, satellite, and future quantum networking environments while maintaining clear separation between physical models, communication protocols, and higher-level decision-making.

The platform provides physically grounded models of decoy-state BB84, atmospheric transmittance, orbital geometry, and channel dynamics while exposing interfaces through which machine learning algorithms can observe, evaluate, and adapt communication behaviour under changing conditions. Rather than serving solely as a quantum key distribution simulator, Quantum-QKD-Aero is designed as an extensible laboratory for exploring adaptive communication strategies across multiple layers of future quantum communication architectures.

As a representative example, the poster outlines an ongoing reinforcement learning study in which an adaptive agent learns communication decisions under time-varying channel conditions. This initial experiment illustrates how the platform supports investigation of adaptive protocol configuration while establishing a foundation for future studies in routing, resilience, anomaly detection, heterogeneous network coordination, and communication across emerging quantum networking paradigms.

The long-term vision is to provide an open, extensible research environment that enables the quantum communication community to investigate adaptive communication as future quantum networks grow in scale, diversity, and operational complexity.

Data-enhanced self-learning projective quantum Monte Carlo with Rydberg-atom measurements

We introduce a data-enhanced self-learning projective quantum Monte Carlo (PQMC) method that incorporates projective measurements from programmable quantum simulators into the guiding wave function of a ground-state projection. We demonstrate the approach for a two-dimensional Rydberg-atom array using restricted Boltzmann machine (RBM) and recurrent neural network (RNN) guiding wave functions. Experimental pretraining reduces finite-population effects and accelerates self-learning, allowing accurate ground-state properties to be obtained with fewer walkers. The improvement is particularly pronounced for the RBM, while the more expressive RNN already provides an accurate guide without pretraining and consequently exhibits a smaller additional gain. Similar improvements are observed for the checkerboard order parameter, despite deviations of the experimental measurements from the equilibrium numerical results. Finally, finite-size scaling with self-learning PQMC without experimental pretraining yields the critical point and critical exponents of the disordered-to-checkerboard transition in agreement with previous numerical studies. Our results show that limited and imperfect experimental measurements can provide an effective initialization for self-learning PQMC, while imaginary-time projection subsequently refines the guiding state toward the ground state.