Preparing quantum states describing systems at finite temperatures presents a major hurdle for modern computations that affect modeling of complex physical systems and advances in quantum machine learning. We address this challenge by introducing a new metal learning algorithm, metavariation thermalizer, and neural network metaVQT, designed to efficiently prepare these thermal states on current quantum computers. These methods learn from a range of system parameters and allow for rapid and accurate thermal conditions to be generated for systems that have never been encountered before. Importantly, the team is introducing practical applications of these algorithms by training quantum Boltzmann machines by significantly faster and faster than existing methods, paving the way for more efficient and scalable quantum machine learning applications.
Inspired by previous research, researchers have developed two new metal learning algorithms, Metavariation Quantum Thermizer (META-VQT) and the neural network Meta-VQT (NN-Meta VQT) to efficiently prepare thermal states on current quantum computers. Both algorithms leverage collective optimization via training sets to generalize Gibbs state preparation to invisible parameters, demonstrating effectiveness in systems up to eight kitz.
Metropolis algorithm for preparation of thermal states.
This study addresses the challenge of preparing thermal states in short-term quantum computers for many quantum algorithms. The authors introduce Meta-VQT and Nn-Meta VQT. It aims to learn how to efficiently prepare these states using variational quantum circuits. The main innovation is the meta-learning approach, where circuit parameters are optimized to minimize the difference between prepared and target thermal states. Meta-VQT employs a variational quantitative circuit to optimize parameters, minimizing the loss function that measures the distance between the prepared state and the target state.
The NN-Meta VQT enhances this by mapping Hamiltonian parameters to circuit parameters using neural networks, allowing for more flexible and expressive parameterization. The team is also applied to train the QUNTUM BOLTZMANN machine (QBMS) of the NN-Meta VQT, preparing the Gibbs state of the QBM and updating its parameters to model the target probability distribution. The results show that the NN-Meta VQT is more robust than the traditional methods and effectively trains QBMs. This study achieves improved accuracy and robustness in thermal condition preparation, utilizes neural networks to increase expressivity and training efficiency, providing thorough experimental setup and clear pseudo-lattice.
Metal learning optimizes preparation for thermal conditions
Researchers have developed new technologies to prepare the state of Gibbs, important for simulating quantum systems, and have used a wide range of applications in areas such as materials science and drug discovery. These methods, called metavariational thermizers (Meta-VQT) and neural networks Meta-VQT (NN-Meta VQT), address the fundamental challenges of quantum computing by efficiently preparing these states in current noisy quantum hardware. Co-innovation lies in a collective optimization approach where algorithms learn to prepare Gibbs across different values as well as specific parameter settings, improving the ability to generalize beyond initial training data. The team's approach is based on existing variational quantum algorithms, but introduces meta-learning components similar to “learning” in classical machine learning.
Instead of training quantum circuits for each new set of parameters that define quantum systems, ALGORITHMS trains systems that can quickly adapt to invisible parameters. This is achieved by training the algorithm on a set of parameters, applying its learned knowledge, and applying the Gibbs state to prepare it for new, previously seen parameters. Demonstrations on systems with up to eight qubits, including transverse magnetic field ISING and HEISENBERG models, show that meta-trained algorithms can accurately generate thermal states beyond the data used for training. For larger systems, the algorithm serves as an effective starting point for optimization, significantly outweighing random initialization. For the 3 quit kitaev ring model, the algorithm demonstrated the ability to effectively capture the operation of the system over different temperature ranges and handle finite temperature effects. The team observed a 30x faster speed compared to existing technologies, improving the accuracy of Gibbs' preparation required for training, suggesting a pathway to scalable and efficient quantum machine learning.
Meta-algorithms generalize preparation for thermal states
The researchers have developed two new meta-algorithms: Metavariation Thermizer (Meta-VQT) and Neural Network Meta-VQT (NN-Meta VQT), designed to efficiently prepare thermal states on short-term quantum computers. These methods move beyond traditional approaches by training the algorithm to learn Gibbs states with different Hamiltonian parameters rather than optimizing each parameter individually. This collective optimization allows generalization to allow for invisible parameter configurations, reduce the required quantum resources and accelerate the process of finding thermal states. Demonstrations on systems with up to eight qubits, including transverse magnetic field ISING and HEISENBERG models, show that meta-trained algorithms can accurately generate thermal states beyond the data used for training.
For larger systems, the algorithm serves as an effective starting point for optimization, significantly outweighing random initialization. Furthermore, this approach successfully prepares Gibbs states across the finite temperature crossover regime, as demonstrated in the Kitaeffling model, strengthens training of the Boltzmann quantum machine, and achieves a 30x runtime speedup compared to existing techniques. The authors acknowledge that their numerical results are currently based on a 2 quit Hamiltonian, and scaling to larger systems and more complex models may require deeper quantum circuits. This study highlights the possibility that meta-algorithms may overcome the limitations of preparing thermal states in noisy, intermediate-scale quantum devices.
