NQS Achieves Competitive Performance With New AGD Optimization

Researchers from Chinese Academy of Sciences have developed a new optimization technique that overcomes a fundamental instability in neural quantum state (NQS) calculations, enabling compact machine learning models to achieve competitive accuracy across a range of challenging quantum systems. Shiwei Zhou and colleagues identified a previously unrecognized finite-sample instability, termed subspace trapping, and introduced annealed gradient descent (AGD) to overcome it. Their results show that the method enables neural quantum states to reach chemical accuracy while improving performance in molecular simulations and quantum many-body models.

Neural quantum states use neural networks to approximate the wavefunctions of quantum systems, offering a powerful alternative to traditional computational methods. However, the researchers found that finite sampling during optimization can produce a self-reinforcing failure mode. In subspace trapping, physically important configurations are assigned probabilities that are too low, causing them to disappear from subsequent sampling batches. Because these configurations are no longer sampled, they receive little or no gradient feedback, preventing the model from recovering them and causing the optimization to converge to an apparently stable state with an energy above the true ground state.

To address this problem, the team developed annealed gradient descent, an optimization strategy that temporarily increases the contribution of low-probability configurations while limiting the influence of highly probable ones during training. This rebalancing helps preserve physically relevant configurations throughout the optimization process, preventing the loss of sampled support that drives subspace trapping.

The researchers evaluated AGD across a variety of benchmark problems, including molecular systems as well as one- and two-dimensional J1-J2 spin models. In each case, the method significantly improved optimization performance, allowing relatively compact neural quantum states to achieve chemical accuracy and remain competitive with more complex approaches. Because AGD operates as an optimization strategy rather than requiring changes to the underlying neural network architecture, it can be readily incorporated into existing neural quantum state workflows.

Beyond improving accuracy, the work provides new insight into why neural quantum state optimization can fail despite appearing to converge successfully. By identifying subspace trapping as a distinct finite-sample instability and demonstrating an effective way to overcome it, the researchers establish a practical framework for more reliable quantum many-body simulations.

As quantum computing and quantum simulation continue to advance, optimization methods that are both efficient and robust will become increasingly important. The lightweight nature of annealed gradient descent makes it a promising addition to existing neural quantum state techniques, with potential applications in quantum chemistry, condensed matt