James WT Keeble and colleagues at TIFPA’s Trent Institute have constructed a neural quantum state representation of an atomic nucleus that is highly entangled and deviates from easily representable “stabilizer” states. These representations show that states exhibiting greater instability are more difficult for the network to learn, and this property has a significant impact on the efficiency of compression and representation using restricted Boltzmann machines. This discovery is key to optimizing network architectures and improving our ability to model highly entangled quantum systems.
Quantum complexity limits neural network modeling of atomic nuclei
The accuracy for representing intermediate-mass nuclei decreased to less than 13% error in the most complex case. This is a significant improvement over traditional methods, which cannot model such highly intertwined systems. This represents a major advance since traditional methods such as exact diagonalization and coupled cluster theory have difficulty calculating for nuclei above a certain mass number due to the exponential scaling of Hilbert space. Representation ability is directly related to ‘astableness’, a measure of quantum complexity, and it has been consistently proven that states that exhibit greater asstability are more difficult to train restricted Boltzmann machines (RBMs). Unstabilizability quantifies the extent to which a quantum state cannot be efficiently described by a stabilizer code, a type of quantum error correcting code. While stabilized states are relatively easy to simulate for classical computers, non-stabilized states require exponentially more resources. While accurate simulations of nuclei exhibiting significant entanglement were previously hampered, there now exists a clear threshold at which traditional neural network approaches have difficulty.
Researchers from the University of Surrey and the University of Edinburgh have tailored a second quantization formulation of neural quantum states for nuclear physics, avoiding computational limitations encountered in previous work and allowing calculations within a manageable parameter space. This formulation exploits the fermion-like nature of the nucleons in the nucleus and uses creation and annihilation operators to describe their behavior. Previous attempts often relied on approaches that were quantized first, which resulted in large computational overheads when dealing with many-body systems. Although detailed analysis reveals a strong correlation between accuracy and “instability,” simple measures of entanglement such as entanglement entropy do not show similar correlations, suggesting that this is a major factor limiting the ability of these networks to compress and represent highly entangled states. Although entanglement entropy indicates overall entanglement, it does not fully capture the specific types of entanglement that impede neural network representation. The researchers found that destabilization is a more nuanced predictor of the difficulties faced by RBMs. Simulations using this approach were able to model the largest nucleus studied, 28 silicon, using about 10% of the parameters required by traditional methods.
Central to this approach is the use of restricted Boltzmann machines (RBMs). RBM is a type of generative probabilistic artificial neural network that can learn complex probability distributions. In this context, RBM learns to represent the wavefunction of the ground state of an atomic nucleus. The parameters of the network are tuned through the training process to minimize the difference between the RBM-generated wavefunction and the true ground state obtained from more accurate but computationally expensive methods. The success of this method depends on the RBM’s ability to efficiently encode the correlations present in the nuclear wavefunction. Although this new approach has successfully mapped the complex behavior of atomic nuclei using neural networks, important limitations remain regarding the computational demands of modeling increasingly complex systems. The accuracy of these networks is strongly influenced by their “nonfactorizability,” or the extent to which the behavior of the nuclei deviates from a simple, predictable pattern. For highly factorizable systems, the wavefunction can be expressed as a product of single-particle states, which greatly simplifies calculations. However, real atomic nuclei exhibit strong correlations between nucleons, leading to significant unfactorization. Addressing these challenges will require significantly more computational power and sophisticated network architectures, and may require exploration of alternative neural network designs and hybrid computational strategies.
The importance of this initiative goes beyond simply achieving reductions in computational costs. This provides fundamental insight into the relationship between quantum complexity and machine learning. Understanding the properties of quantum states that are most difficult for neural networks to represent is important for developing more effective algorithms for quantum simulations. This study establishes a valuable baseline for the use of artificial intelligence in nuclear physics and recognizes the great challenge of representing highly complex atomic nuclei. A direct relationship exists between the complexity of a quantum state and the ability of a restricted Boltzmann machine (RBM) to accurately represent the quantum state. When modeling intermediate-mass nuclei, we found that higher degrees of “destabilization” become more difficult to learn systematically, providing important insights for refining these networks. This finding highlights that RBM is a key factor in determining how efficiently entangled quantum systems can be compressed and represented, and will initiate further developments to enable even more complex simulations. The ability to accurately model larger, more complex atomic nuclei could improve our understanding of nuclear structure, nuclear reactions, and the synthesis of heavy elements, and expand the applicability of neural networks to previously inaccessible areas of nuclear physics. Future research may focus on exploring different network architectures, such as variational quantum circuits, and developing more efficient training algorithms to overcome the limitations imposed by astableness and nonfactorizability.
This study demonstrated that restricted Boltzmann machines (RBMs) struggle to accurately learn representations of quantum states with higher “instability.” This is important because it identifies important properties of quantum systems that influence how efficiently neural networks can compress and represent complex information. This finding suggests that astabilization is the main factor limiting the performance of RBMs when modeling entangled systems such as intermediate-mass nuclei. The authors suggest that future research explore more sophisticated network architectures to address these limitations and improve representation efficiency.
