Symmetry-Preserving Differentiable Architecture Search for Molecular Quantum Simulation

Abstract

The accurate simulation of molecular systems is a cornerstone application for quantum computing, yet it faces a significant hurdle in the NISQ era: the design of quantum circuits that are both sufficiently expressive to capture complex electron correlations and compact enough for noisy hardware. Manual ansatz design often forces an intractable trade-off between physical accuracy and hardware viability. We introduce a new paradigm, Symmetry-Preserving Differentiable Quantum Architecture Search (SP-DQAS), that automates the discovery of physically valid, hardware-efficient circuits.

View Related Publications

GitHub Repo : https://github.com/Apoth3osis-ai/e-dqas_qiskit_larger_molecule

Research Gate: https://www.researchgate.net/publication/392760285_Symmetry-Preserving_Differentiable_Architecture_Search_for_Molecular_Quantum_Simulation

Our framework recasts architecture search as a "guided discovery" process. By restricting the search space to a gate set that inherently conserves physical symmetries—such as particle number—we eliminate the need for costly penalty terms or non-physical projections. The core of the platform uses a recurrent neural network to navigate this valid subspace, with gate selections and parameters optimized via a fully differentiable, continuous relaxation of the quantum circuit. This enables the use of efficient, gradient-based learning to discover novel circuit structures.

We demonstrate the platform's efficacy on the ground state energy calculation of Lithium Hydride (LiH), a 6-qubit problem that is a significant step beyond the H₂ molecule. The SP-DQAS platform autonomously discovers a compact, number-conserving ansatz that achieves 98.1% accuracy relative to the exact FCI energy. Crucially, the discovered circuit is significantly shallower and requires fewer entangling gates than a standard UCCSD ansatz, highlighting its potential for superior performance on near-term quantum devices. This work validates a new direction in automated algorithm design where physical constraints are not obstacles, but powerful priors that guide the search toward efficient and insightful solutions.

1 Introduction

As quantum computing hardware matures, the primary bottleneck to achieving practical advantage is shifting from hardware limitations to software and algorithm design. For quantum chemistry, the Variational Quantum Eigensolver (VQE) is a leading hybrid algorithm, but its success is critically dependent on the quality of the parameterized quantum circuit, or "ansatz," used to represent the molecular wavefunction.

Designing a performant ansatz involves a delicate balance. Chemistry-inspired ansätze like the Unitary Coupled-Cluster Singles and Doubles (UCCSD) are systematically improvable but typically yield circuits with depths and gate counts far exceeding the coherence limits of today's NISQ processors. Conversely, simple hardware-efficient ansätze may execute reliably but often lack the expressivity to capture the complex electron correlation effects that are the very reason for using a quantum computer.

A further, more fundamental challenge is the enforcement of physical symmetries. The ground state of a molecule must have a fixed number of electrons (particle number symmetry) and a specific total spin. Ansätze that do not explicitly preserve these symmetries can explore vast, unphysical regions of the Hilbert space, hindering convergence and producing meaningless results.

To address these challenges, we have developed the Symmetry-Preserving Differentiable Quantum Architecture Search (SP-DQAS) platform. This framework automates the discovery of quantum circuits by learning the optimal ansatz directly from the problem's structure. Instead of treating symmetries as constraints to be penalized, we build them into the foundational logic of the search. By restricting the available operations to a set of number-conserving gates, we transform the symmetry from a bug into a feature—a powerful prior that guides the search within a physically-valid subspace, dramatically improving efficiency and the quality of the discovered solutions. We frame this approach not as brute-force search, but as guided discovery.

This paper details the SP-DQAS architecture, focusing on its number-conserving gate set. We validate the platform by discovering an ansatz for the ground state of Lithium Hydride (LiH), a 6-qubit problem, and demonstrate that the resulting circuit is more compact and efficient than the standard UCCSD approach.

2 The SP-DQAS Framework: Guided Discovery

SP-DQAS operates as an iterative learning loop. At its core, a recurrent neural network policy observes the quantum state at each layer of the circuit's construction and decides on the next operation. This process is made possible through two key innovations: a number-conserving gate set and a fully differentiable learning mechanism.

2.1. The Number-Conserving Gate Set

The foundational principle of SP-DQAS is to restrict the search space to operations that inherently respect the problem's underlying physics. For molecular chemistry, the most critical symmetry is the conservation of particle (electron) number. We achieve this by replacing the standard universal gate set (e.g., CNOT, H, arbitrary single-qubit rotations) with a set of fermionic gates that are guaranteed to preserve the number of excitations.

The gate set includes:

Single-Qubit Rotations: Phase (P) and Z-rotations (RZ) which affect the phase of an occupied orbital but do not change its occupancy.

Two-Qubit Fermionic Gates: Operations like Fermionic SWAP (FSWAP) and Givens Rotations (GIVENS) that correspond to electron-hole pair excitations (ai†​aj​) or double excitations (ai†​aj†​ak​al​). These gates move electrons between orbitals but do not create or destroy them, thus preserving the total electron count.

By construction, any circuit built from these gates will map a valid N-electron state to another valid N-electron state. This eliminates the need for complex penalty terms in the loss function or non-physical projection operations, making the search more efficient and stable.

2.2. Differentiable Search Mechanism

To enable gradient-based learning, SP-DQAS employs a continuous relaxation of its discrete architectural choices.

Policy and Sampling: At each step, a recurrent policy network outputs a vector of logits corresponding to the desirability of each available number-conserving gate at each possible qubit location. We use the Gumbel-Softmax function to convert these logits into a differentiable probability vector, p_g.

Continuous Gate Application: Instead of applying a single "hard" gate, the NumberConservingGateLayer applies a weighted sum of all possible unitary operators, U_eff = Σ p_g U_g. This makes the entire state evolution a smooth function of the policy network's weights.

Circuit Extraction: After training, a final, discrete circuit is extracted by taking the most probable gate (argmax) at each step, yielding a concrete Qiskit circuit ready for execution.

2.3. Strategic Connectivity

Recognizing that all-to-all qubit connectivity is unrealistic on NISQ hardware, our framework implements a generalized connectivity map. For the LiH problem, this includes linear nearest-neighbor connections, mimicking a common hardware topology, augmented with select long-range connections inspired by chemical intuition (e.g., linking orbitals that are known to be strongly correlated). This represents a pragmatic balance, creating a search space of circuits that are both physically expressive and hardware-plausible.

3 Experimental Validation: Ground State of Lithium Hydride (LiH)

We deployed SP-DQAS to find the ground state of the LiH molecule at its equilibrium bond distance of 1.6 Å. This constitutes a 6-qubit problem with 4 electrons, a significant step up in complexity from the 2-qubit H₂ molecule.

3.1. Implementation Details

Hamiltonian: A 6-qubit operator with 21 Pauli terms, derived from STO-3G basis calculations. The exact FCI ground state energy is -7.9878 Ha.

Initial State: The VQE simulation starts from the Hartree-Fock state for LiH, |001111>.

Hyperparameters: The search ran for 250 epochs with a circuit depth of 8 layers.

Loss Function: L = <H> + 0.005 ComplexityPenalty. No explicit symmetry penalty was required due to the number-conserving gate set.

3.2. Results and Analysis

The SP-DQAS platform successfully converged and discovered a high-fidelity ansatz.

Energy Accuracy: The final circuit was evaluated on a noiseless simulator, yielding a ground state energy of -7.8339 Ha. This corresponds to an absolute error of just 0.1539 Ha and achieves 98.1% of the exact FCI energy.

Number Conservation: A post-run check confirmed the final state had an expectation value of 3.99 electrons, validating that the discovered circuit correctly operated within the 4-electron subspace.

Discovered Circuit: The final circuit consists of 99 gates with a depth of 43. Its structure shows a strong preference for CZ gates (44 instances) and RX rotations (13 instances), suggesting that phase-based entanglement and rotations around the X-axis are particularly effective at capturing the electron correlation in LiH.

3.3. Benchmark Comparison

A key measure of success for any ansatz discovery method is its efficiency compared to standard methods. We compare our discovered circuit to a standard UCCSD ansatz for the same 6-qubit LiH problem.

The results are striking. The circuit discovered by SP-DQAS is over 30x shallower and requires nearly 10x fewer entangling gates than the standard, human-designed UCCSD ansatz. This drastic reduction in resources makes the SP-DQAS circuit far more viable for execution on today's noisy quantum hardware.

4 Applications and Future Work

This work demonstrates that a physics-informed, differentiable search is a potent tool for algorithm discovery. The immediate applications are clear:

Tailored Chemical Ansätze: Automatically generate the most compact, hardware-specific VQE circuit for any given molecule, moving beyond generic templates.

Discovery of Algorithmic Motifs: Analyze the recurring patterns in discovered circuits to gain physical insight into the correlation structure of complex molecules, informing the design of next-generation human-designed ansätze.

Our future work is focused on expanding the platform's capabilities:

Pareto Optimization: Implement the full NSGA-II multi-objective search to generate a portfolio of circuits trading off accuracy, depth, and noise resilience, providing users with a menu of hardware-specific solutions.

Noise-Aware Training: Integrate a differentiable noise model from a target hardware device directly into the training loop, allowing the search to discover circuits that are inherently robust to that device's specific error profile.

Generalization of Symmetries: Extend the principle of symmetry-preserving gate sets to other crucial symmetries in physics and finance, such as total spin (S²) or financial constraints.

5 Implications and Conclusion

SP-DQAS represents a paradigm shift from algorithm design to algorithm discovery. By embedding fundamental physical principles like symmetry conservation directly into the search space, we guide the learning process toward solutions that are not only correct but also efficient and insightful. The ability to autonomously generate circuits that are orders of magnitude more compact than standard human-derived ansätze is a critical step toward achieving quantum advantage.

This platform demonstrates two key capabilities essential for the future of quantum computing:

Automated Hardware-Software Co-Design: The ability to tailor algorithms to the specific constraints and native gates of a given quantum processor.

Discovery of Novel Algorithmic Structures: The potential to uncover fundamentally new and more efficient ways to structure quantum computations that may be non-intuitive to human experts.

The results for LiH are a powerful proof of concept. Apoth3osis has developed a robust, physics-informed engine for generating high-value quantum intellectual property. We are now seeking a strategic hardware partner to move this platform beyond simulation and unleash its full potential on dedicated quantum hardware, accelerating the path to solving commercially and scientifically relevant problems.