AI Ranks Tensor Network Plans for Faster Quantum Simulation

Classical simulation is a vital tool for developing and validating quantum algorithms, but its computational cost escalates rapidly with the complexity of quantum circuits. Tensor-network contraction offers a way to mitigate this cost by exploiting the inherent structure of circuits. However, the efficiency of this method heavily depends on the specific contraction plan chosen, as plans with similar theoretical complexity can perform very differently on GPUs due to factors like parallelism, memory traffic, and contraction geometry. To address this, researchers have developed a learning-to-rank framework designed to select optimal contraction plans before their execution. Each potential plan is characterized by structural features derived from its sequence of pairwise contractions. Gradient-boosted rankers are then trained using GPU performance measurements, employing both listwise and pairwise objectives. The evaluation of these models across diverse circuit families, including tests for distribution shifts, showed that the learned rankers consistently identify superior plans compared to random or MinFill-based baselines, with the listwise model demonstrating the highest overall decision quality. The study also examined backend shift, finding that plan rankings remained largely stable across different GPU architectures, indicating that models retain useful decision quality even without retraining. This framework offers a practical solution for reducing the search space for efficient contraction plans in quantum circuit simulation.