QAELL: Toward Quantum-Assisted Energy-Landscape Learning for Anomaly Detection

Energy-based semi-supervised anomaly detection learns to assign lower energy to nominal observations than to scarce labeled anomalies. Yet finite supervision underconstrains the energy landscape, allowing unsupported configurations to receive nominal-like energy. We introduce Quantum-Assisted Energy-Landscape Learning (QAELL), a variational autoencoder with a restricted Boltzmann machine (RBM) prior that combines nominal reconstruction and prior fitting, supervised energy separation, and low-energy state discovery. A coherent Ising machine (CIM) is an Ising optimizer whose finite-run reads favor low-energy configurations; QAELL maps its latent prior to Ising form and uses these states in a detached negative phase, while inference remains classical. A one-step analysis links overrepresentation by the negative phase to energy-raising prior updates. A fully enumerable planted-basin study demonstrates repair of a constructed unsupported region under finite training updates. In a credit-card fraud case study, QAELL-CIM achieves competitive average precision, close to its classical-sampler counterparts and above the tested unsupervised and semi-supervised baselines. This demonstrates the feasibility of integrating a CIM negative phase into an anomaly detector with classical inference.