Designing a classical neural network architecture is challenging, but engineers can rely on strong spatial and structural intuitions—such as using convolutional layers to process images or recurrent blocks for sequential data. Quantum circuit design enjoys no such luxury.
In a variational quantum circuit (VQC) or parameterized quantum circuit (PQC), a designer must decide exactly where to place single-qubit rotations and multi-qubit entangling gates. As noted by Yuxuan Du et al. in Quantum circuit architecture search for variational quantum algorithms, this manual process is highly heuristic and lacks theoretical guarantees. Human intuition often defaults to rigid, repeating templates (ansatzes) that fail to match the specific mathematical structure of a problem or the physical constraints of noisy intermediate-scale quantum (NISQ) hardware.
To break this bottleneck, Quantum Architecture Search (QAS) treats the topology of a quantum circuit not as a fixed human guess, but as a discrete optimization problem. By automating the discovery of gate layouts, QAS uncovers non-intuitive, hardware-efficient configurations that outperform traditional human designs.
The Discrete Search Space Problem
Why is finding a good quantum circuit so difficult? A quantum circuit is essentially a directed acyclic graph of discrete gate operations chosen from a native hardware library (such as CNOT, $R_x$, $R_y$, and $R_z$). The number of possible configurations grows exponentially with the number of qubits and circuit depth. Furthermore, as Johannes Jakob Meyer et al. point out in Exploiting symmetry in variational quantum machine learning, arbitrary gate choices often fail to respect the underlying symmetries of the data, leading to poor generalization or severe optimization hurdles like barren plateaus.
Early attempts to automate this process relied on evolutionary and genetic algorithms. The foundational work by Colin P. Williams et al. in Automated Design of Quantum Circuits demonstrated that heuristic search could discover novel gate configurations that human intuition overlooked.
Modern evolutionary frameworks, such as the one proposed by Li Ding et al. in Evolutionary Quantum Architecture Search for Parametrized Quantum Circuits, have shown that automated search can discover highly counter-intuitive but performant blueprints—such as placing data-encoding operations exclusively at the front of a reinforcement learning circuit, mimicking the input layer of a classical neural network.
Reinforcement Learning and Adaptive Exploration
To navigate this vast discrete space more efficiently than random mutation, researchers use deep reinforcement learning (RL). In this framing, an RL agent acts as the architect, sequentially selecting gates to construct a circuit. The "reward" given to the agent is tied to the performance of the trained circuit on a target task.
The power of this approach lies in its ability to discover task-specific shortcuts. For instance, Xin Dai et al. developed an adaptive search strategy in Quantum Machine Learning Architecture Search via Deep Reinforcement Learning that shifts the agent's optimization goals as it learns, successfully decoupling model design from expert templates.
Similarly, Tuomas Laakkonen et al. leveraged tensor decomposition techniques within an RL agent in Quantum circuit optimization with AlphaTensor to find non-obvious, human-transcending paths to minimize computationally expensive T-gates in fault-tolerant regimes.
Differentiable Search: The Continuous Relaxation
While RL and evolutionary algorithms are powerful, they treat circuit evaluation as a black box. They require training thousands of candidate circuits from scratch to evaluate their fitness, which is computationally expensive.
This challenge inspired Differentiable Quantum Architecture Search (DQAS), introduced by Chang-Yu Hsieh et al. in Differentiable quantum architecture search. DQAS works by relaxing the discrete choice of gates into a continuous, probabilistic mixture, allowing researchers to use gradient descent to optimize the structure of the circuit alongside its parameters.
Imagine a specific slot in a circuit that could contain an $R_x$ gate, a CNOT gate, or an identity operation (no gate). Instead of picking one, DQAS assigns a continuous weight parameter $\alpha_i$ to each candidate gate $G_i$. The operation at that slot becomes a superposition of gates:
$$U = \sum_i \sigma(\alpha_i) G_i$$
where $\sigma$ is a softmax function that normalizes the weights. Because the entire structure is now parameterized by continuous weights $\alpha$, a classical computer can compute gradients with respect to both the gate parameters (such as rotation angles $\theta$) and the structural parameters $\alpha$ simultaneously. Once optimization converges, the highest-weighted gates are selected, collapsing the continuous mixture back into a discrete quantum circuit.
Discrete Choice (Hard to optimize) Continuous Relaxation (Differentiable)
[ Slot 1 ] [ Slot 1 ]
/ | \ / | \
[Rx] [CNOT] [I] α₁[Rx] α₂[CNOT] α₃[I]
\ | /
U = Σ σ(α_i) G_i (Gradient Descent)
This approach drastically accelerates search efficiency. Yize Sun et al. applied this method to reinforcement learning in Differentiable Quantum Architecture Search for Quantum Reinforcement Learning, discovering highly unusual gate arrangements—like embedding entangling operations deep within parameterized blocks rather than at boundaries—that outperformed human templates by 200%.
The authors extended this to combinatorial optimization in Differentiable Quantum Architecture Search For Job Shop Scheduling Problem, demonstrating that DQAS naturally tailors architectures to the specific mathematical structure of a scheduling problem while maintaining noise resilience.
Hardware-Aware and Multi-Objective Constraints
An optimal circuit is not just one that achieves high accuracy; it must also survive the noise of real-world quantum processors. Every additional gate introduces decoherence and operational errors. QAS frameworks are uniquely suited to handle this because they can optimize for multiple competing objectives simultaneously.
In Multi-Objective Evolutionary Architecture Search for Parameterized Quantum Circuits, Li Ding 0010 et al. demonstrated a framework that explicitly penalizes circuit depth and gate counts alongside task error. By doing so, the search algorithm naturally prunes away noise-sensitive configurations, finding compact, hardware-efficient topologies.
This hardware-aware co-design is critical when dealing with physical constraints. As shown by Tim Menke et al. in Automated design of superconducting circuits and its application to 4-local couplers, an automated closed-loop system successfully engineered physical superconducting couplers that surpassed human-designed proposals in noise resilience and coupling strength.
By shifting the burden of design from human trial-and-error to algorithmic optimization, QAS provides a scalable path toward practical quantum advantage on both near-term and fault-tolerant hardware.
Key Foundational Work
Automated Design of Quantum Circuits (1998) — contributed 1 concept elements to this direction.
Quantum algorithms for the ordered search problem via semidefinite programming (2007) — contributed 1 concept elements to this direction.