Quantum Model-Discovery (QMoD)

• Quantum Model-Discovery (QMoD) is a framework that uses quantum devices and hybrid quantum–classical methods to reconstruct underlying quantum models, dynamics, and constraints from data.
• It enables causal and structural discovery through quantum tomography and kernel-based algorithms, providing efficient techniques for identifying unique quantum process parameters.
• The approach also extends to scientific machine learning by automating the discovery of dynamical equations and generating quantum materials via constrained diffusion models.

Quantum Model-Discovery (QMoD) denotes a broad methodological and conceptual framework in which quantum devices, hybrid quantum–classical algorithms, or quantum-informed generative models are actively used to discover, from data or black-box access, the underlying structure, dynamics, or constraints of quantum and quantum-inspired systems. Applications extend across quantum device tomography, causal structure inference, learning of dynamical equations, the discovery of new quantum materials, and the automated search for nontrivial many-body dynamics. The essential feature is the formulation and solution of an inverse problem in which the quantum theory model—parameters, operators, or dynamical rules—are inferred or reconstructed from experimental or numerical data, often exploiting quantum computational or representational resources.

1. Foundational Principles and Theoretical Guarantees

At its core, QMoD is concerned with model reconstruction: Given empirical statistics or outcomes from a quantum device or process, determine the underlying model (set of state preparations, quantum channels, measurement operators, or structural rules) consistent with quantum theory and, ideally, uniquely specified up to physically trivial equivalence classes.

A canonical formulation (Ren et al., 2019) uses the triplet of experimental controls: a set of prepared states $\{\rho_i\}$, a collection of quantum channels or maps $\{\mathcal{M}_j\}$, and a measurement ensemble $\{E_k\}$. The observable statistics

$$p_{\mathfrak{m}}(i, j_1, \dots, j_N, k) = \mathrm{Tr}\left[E_k\, \mathcal{M}_{j_N} \circ \dots \circ \mathcal{M}_{j_1}(\rho_i)\right]$$

define a "representation" $\mathfrak{m}$ of a physical model $M$. The essential problem is to invert $$p_{\mathfrak{m}}\mapsto \{\rho_i, \mathcal{M}_j, E_k\}$$, modulo unavoidable Wigner-type symmetries (unitary and anti-unitary conjugations), and to settle the conditions for uniqueness.

Uniqueness Theorems:

Statistical indistinguishability can arise from nontrivial gauge freedoms, captured by invertible, Hermitian- and trace-preserving maps $\mathcal{T}$: $\mathcal{T}(X) = \sum_l \eta_l F_l X F_l^\dag, \quad \eta_l\in\mathbb{R},\; \sum_l \eta_l F_l^\dag F_l = \openone.$ Necessary and sufficient conditions were established (Ren et al., 2019):

• Necessary: Nontrivial uniqueness requires (at least) one singular state or channel (i.e., $\prod_i \det(\rho_i)\prod_j \det(C_j) = 0$ for Choi states $C_j$).
• Sufficient (i): If the preparations and measurements span a complete rank-1 projective ensemble $\Pi$ (computational basis, two- and three-state superpositions), any physically valid model is unique up to global Wigner-type symmetries.
• Sufficient (ii): If the device implements the full unitary group and at least one state and one POVM element are rank-deficient, uniqueness is again guaranteed.

The practical implication is that generic quantum devices subject to restricted controls are generically non-identifiable, unless high-quality tomographic completeness or symmetry-generating ensembles are engineered.

2. Quantum-Enhanced Causal and Structural Discovery

QMoD encompasses a class of quantum-enhanced algorithms for causal inference, structure recovery, and process tomography in networks and dynamical systems. Approaches span direct quantum causal structure algorithms, quantum-kernel-driven classical inference, and efficient order-discovery in quantum networks.

Quantum Causal Model Discovery:

• Process-matrix-based algorithms (Giarmatzi et al., 2017) recover, from an $n$-party process matrix $W$, (i) whether the system is causally ordered, (ii) whether it is Markovian for a given DAG (via specific factorization properties), and (iii) the minimal DAG and the quantum mechanisms generating $W$. The algorithm proceeds in stages: identifying maximal non-signaling sets, extracting the DAG (by tracing open outputs), and confirming Markovianity by reconstructing $W$ from the inferred mechanisms. Computational complexity is nominally $O(n^2)$ tests, limited by the exponential growth of Hilbert space but tractable for moderately sized processes.
• Quantum comb approaches (Bai et al., 2020) generalize the above to causal order-discovery in arbitrary quantum circuits. Recursive identification of "last" input-output pairs, followed by sequential removal, enables polynomial-scaling discovery of causal orderings, provided the quantum comb has low Kraus rank. Under more restrictive structural assumptions (total influence or memoryless channels), purely local observations suffice for polynomial recovery of the causal order.

Quantum-enhanced kernel methods:

• Quantum RKHS causal discovery (Terada et al., 9 Jan 2025, Kawaguchi, 2021): Quantum circuits parameterize feature maps $\Phi(x)=U_\theta(x)|0^n\rangle$, inducing highly expressive kernels $k_Q(x,x';\theta)$ for kernel-based conditional independence testing (e.g., HSIC). The qPC algorithm generalizes the classical PC algorithm, offering substantial sample efficiency improvements in causal-structure inference for small $n$. Regularization and kernel hyperparameters are data-adaptively chosen via kernel target alignment (KTA), optimizing the circuit for low type-I error in independence testing.

Empirical findings across synthetic and real-world datasets (including biomedical) demonstrate that quantum-enhanced causal discovery yields more accurate DAG reconstruction than classical kernels, especially when available data is limited (Terada et al., 9 Jan 2025, Kawaguchi, 2021).

3. Quantum Model-Discovery in Scientific Machine Learning

A distinct branch of QMoD addresses the automated discovery of dynamical equations from data using quantum circuits as function approximators (Heim et al., 2021). Here, the approach blends ideas from sparse regression (e.g., SINDy, DeepMoD), physics-informed neural networks (PINNs), and quantum variational algorithms.

Equation Discovery Framework:

• The protocol infers a sparse PDE or ODE

$$\partial_t u(x, t) = \sum_{i=1}^P w_i \mathcal{L}_i[u](x, t)$$

where $\{\mathcal{L}_i\}$ are a library of candidate differential operators.

• A differentiable quantum circuit (DQC) with parameters $\theta$ is trained to represent a surrogate solution $\hat{u}(x, t; \theta)$. The loss function combines data fidelity, PDE residual, and $\ell_1$ regularization promoting sparsity in $w$.
• Parameter recovery proceeds by hybrid quantum–classical optimization: quantum hardware estimates outputs and parameter-shift gradients; classical routines update $\theta$ and $w$ jointly.

Proof-of-concept experiments recover parameters and functional form for the damped harmonic oscillator, Lotka–Volterra system, and Burgers’ equation, matching classical regression methods but leveraging the quantum circuit's high-dimensional function space. Resource requirements (qubit count, circuit depth) are modest for low-dimensional problems, although quantum advantage is currently unestablished for large-scale or noisy data (Heim et al., 2021).

4. Generative and Sampling-Based QMoD for Quantum Materials

Generative modeling extended to QMoD is exemplified by the use of structurally constrained diffusion models for quantum material discovery (Okabe et al., 2024). The SCIGEN framework integrates geometric-pattern constraints as hard priors during generative sampling:

• Architectural Details: Crystal structures are parameterized as $(L, F, A)$ (lattice, fractional coordinates, atomic types). A base diffusion model injects Gaussian noise and is trained to denoise; constraints are enforced by masked insertion of prototype structures at every step.
• Masked Conditional Sampling: By diffusing the constraint trajectory and iteratively applying a masking operator, the algorithm samples from the true conditional distribution $p(x_0|c_0)$, ensuring the generated structure exactly incorporates geometric constraints.
• Pipeline: Surviving structures are filtered by multi-stage pre-screening (chemical neutrality, occupancy, GNN classifiers) and validated via high-throughput DFT (with $\sim$95% convergence and $\sim$53% to minima).
• Materials Insights: By steering generation towards e.g., Archimedean lattice motifs, SCIGEN enables the high-throughput discovery of candidate materials supporting flat-band physics, geometric frustration, or topological phenomena, demonstrating QMoD as a composable engine coupling statistical sampling with domain-specific priors (Okabe et al., 2024).

5. Adaptive Discovery Using Interest Functions in Quantum Circuits

A novel variant of QMoD leverages learning agents and quantum hardware to search for "interesting" many-body dynamics by maximizing an explicitly chosen interest function over a circuit family (Placke et al., 1 Jul 2025).

Core Framework:

• The agent optimizes over a circuit family $\{U(\boldsymbol{\theta})\}$ to maximize

$I\big(U(\boldsymbol{\theta})\big)$

where $I(\cdot)$ encodes physical features of interest (e.g., classifiability of states, spectral rigidity).

• The optimization loop alternates quantum evaluation of $I$, via repeated measurement and embedding (e.g., classical shadows, spectral form factor estimation), and classical updates to $\boldsymbol{\theta}$.

Representative Interest Functions:

• Classifiability: Maximized by circuits generating discrete time crystals (DTCs), operationalized as high separability in clustering of evolved state snapshots.
• Spectral Properties: Maximized (or minimized) by dual-unitary circuits, determined by structure in spectral form factor.

Simulations demonstrate efficient recovery of DTC parameters and precise location of dual-unitary regimes in parameter space. The flexibility of the interest function approach suggests systematic, agent-driven exploration of phase diagrams and dynamic phenomena, contingent on quantum hardware scaling and advances in measurement/sampling efficiency (Placke et al., 1 Jul 2025).

6. Limitations, Open Problems, and Experimental Considerations

Several key limitations and open questions typify QMoD research:

• Identifiability: Generic quantum models are not uniquely reconstructible from tomographic data unless stringent conditions (projective completeness, symmetry generation, singularities) are met (Ren et al., 2019).
• Sample Complexity: Quantum-enhanced causal discovery and generative QMoD methods offer sample advantages in practice, but the scaling with system size, effect of noise, and fundamental lower bounds remain active areas of investigation (Terada et al., 9 Jan 2025, Okabe et al., 2024).
• Hardware Limitations: For quantum circuit-based QMoD, performance in realistic settings is constrained by decoherence, finite shot noise, and circuit depth restrictions (e.g., validation on IBMQ for causal discovery with restricted 4–5-qubit circuits) (Kawaguchi, 2021).
• Generality and Expressiveness: The expressivity of DQC ansätze for function approximation, and the sufficiency of operator libraries for equation discovery, define the practical reach of these methods in scientific ML contexts (Heim et al., 2021).
• Interest Function Design: The choice and tuning of interest functions in discovery-mode QMoD are not algorithmically principled; current implementations exploit physics intuition or test a small set of hypothesis-guided observables (Placke et al., 1 Jul 2025).
• Scaling: Routine applications to high-dimensional systems or networks await further algorithmic advances (e.g., quantum eigenvalue estimation for kernel methods) and improved theoretical understanding of quantum inductive bias (Terada et al., 9 Jan 2025).

7. Summary Table: Selected QMoD Paradigms and Contexts

Paradigm	Core Model Objects	Methodology
Full-model tomography (uniqueness)	$\{\rho, \mathcal{M}, E\}$	Analytical gauge analysis
Quantum causal discovery (process matrices)	Process matrix $W$, combs	Non-signaling sets, DAG discovery
Quantum-kernel causal inference	Kernelized independence tests	Parameterized quantum circuits
Quantum equation discovery	Surrogate PDE models	Differentiable quantum circuits
Generative QMoD for materials	Crystal parameterization (L, F, A)	Diffusion with masked constraints
Automated "interest" optimization	Circuit parameters $\boldsymbol\theta$	Interest-function maximization

This table summarizes principal QMoD settings with their main model classes and methodological approach (Ren et al., 2019, Bai et al., 2020, Terada et al., 9 Jan 2025, Heim et al., 2021, Okabe et al., 2024, Placke et al., 1 Jul 2025).

• "Modelling Quantum Devices and the Reconstruction of Physics in Practical Systems" (Ren et al., 2019)
• "A quantum causal discovery algorithm" (Giarmatzi et al., 2017)
• "Efficient Algorithms for Causal Order Discovery in Quantum Networks" (Bai et al., 2020)
• "Quantum-enhanced causal discovery for a small number of samples" (Terada et al., 9 Jan 2025)
• "Application of quantum computing to a linear non-Gaussian acyclic model for novel medical knowledge discovery" (Kawaguchi, 2021)
• "Quantum Model-Discovery" (Heim et al., 2021)
• "Structural Constraint Integration in Generative Model for Discovery of Quantum Material Candidates" (Okabe et al., 2024)
• "Running Quantum Computers in Discovery Mode" (Placke et al., 1 Jul 2025)