Accurate characterization of open many-body quantum systems is constrained by information masking due to exponential growth in Hilbert space and environmental noise. Existing deep learning methods often converge to trivial solutions due to the lack of inductive bias for nonlocal entanglement. To address this, we propose an active learning, entanglement-aware multitask network (AEMTN), a framework designed to efficiently reconstruct disordered Hamiltonians from open dynamics. This architecture overcomes learnability barriers by synergistically integrating three core mechanisms: leveraging entanglement-aware routing to assign task-specific subspace mixtures, applying entanglement-guided inductive biases to impose geometric constraints on the solution space, and leveraging uncertainty modulation of physical correlations to balance gradient flow. Consequently, the framework explicitly translates physical constraints into gradient guidance for inverse parameter identification. Extensive numerical simulations on open quantum spin chains demonstrate the model’s exceptional data efficiency, achieving a dynamical prediction error of $3.0\times {10}^{-4}$ and a correlation coefficient of 99.4%. Crucially, AEMTN demonstrates robust performance in complex phases such as many-body localization and strongly entangled critical points, significantly outperforming baseline models that fail to capture deep nonlocal correlations.

Close

PrevNext

View figure in article

• Machine learning
• Open quantum systems & decoherence
• Quantum information theory