Overview
To automate the feedback control of open quantum systems, this study presents a quantum transformer-based neural network architecture. These models overcome the shortcomings of conventional recurrent neural networks by efficiently processing long-range relationships in measurement data through the use of self-attention processes. The study shows that this method effectively controls complicated non-Markovian dynamics where previous actions have a major impact on future states and stabilizes two-level systems.
Additionally, by utilizing reinforcement learning, the design demonstrates scalability to many-body systems, including Ising chains. The quantum transformer does inference orders of magnitude quicker than traditional numerical solvers, providing notable computational speedups in addition to its adaptability. In the end, this framework offers a reliable and effective way to precisely manipulate quantum technologies in noisy settings.
The Limits of Traditional AI in the Quantum Realm
Researchers have been looking at machine learning for years as a means of producing condensed representations of the correlations seen in quantum data. Although model-free and model-based reinforcement learning have shown potential, earlier designs, such as Recurrent Neural Networks (RNNs), have reached a dead end. RNNs introduce a “Markovian inductive bias” through their step-by-step processing of input. This indicates that individuals have trouble taking into consideration long-range relationships, in which a system’s future is significantly influenced by its remote past.
Because measurement backaction guarantees that future states are inextricably tied to previous measurement outcomes, even seemingly simple systems in the quantum realm can display non-Markovian behavior. Additionally, “vanishing gradients,” a technical flaw that hinders RNNs’ ability to learn from lengthy data sequences, are a common problem.
The Quantum Transformer Enters
Researchers used the quantum transformer architecture, the same technology that drives contemporary natural language processing, to get around these obstacles. In contrast to RNNs, transformers process whole data sequences at once via a method known as self-attention, which enables them to identify correlations between every element of an input sequence.
A bespoke model, including a QuantumEncoder and a QuantumDecoder, was created by the researchers. The measurement record and the initial quantum state are processed by the encoder and then embedded into a high-dimensional space. The ideal control settings are then adaptively determined by the decoder using this compressed representation and positional embeddings of the measurement record. Importantly, the decoder employs “causal masking” to make sure it only examines recent and historical measurements, avoiding “cheating” by viewing data from the future while training.
Performance and Speed: A Two-Order Magnitude Leap
This architecture’s efficacy was initially evaluated on a two-level system (TLS) that was being continuously measured with weak signals. The objective was to use a control parameter, λt, within the Hamiltonian to stabilize the system in a certain target state.
The outcomes were remarkable. In addition to being resistant to measurement errors and system dynamics disturbances, the quantum transformer-based method showed a significant boost in operating speed. Compared to conventional techniques such as the Proportional and Quantum State (PaQS) algorithm, the transformer inferred optimal control settings two orders of magnitude quicker in numerical testing conducted on a typical laptop.
The quantum transformer accomplished the same thing in 0.23 seconds, but the PaQS method takes about 19 seconds per trajectory and necessitates solving intricate stochastic master equations at each time step. Although the researchers pointed out that the significant amount of memory needed to hold the neural network remains a possible constraint for hardware integration, this performance advantage is crucial for real-time feedback.
Controlling Many-Body and Non-Markovian Systems
Applying the transformer to non-Markovian dynamics was the real test of its adaptability. The researchers established a scenario in which the two-level system had “memory” by linking it to a harmonic oscillator mode. The transformer outperformed both vanilla RNNs and Gated Recurrent Units (GRUs) by fine-tuning a pre-trained model through transfer learning. The transformer’s capacity to handle arbitrary lengths of the measurement record allowed it to take precedence as the time horizon increased, even if RNNs performed somewhat better during extremely brief periods of time.
Additionally, the researchers expanded their study to include many-body state preparation, focusing on an N-qubit Ising chain. Since it is frequently hard to discover locally optimum solutions in these complicated systems, there is no “labeled” data to train the AI on. The researchers used model-free reinforcement learning to address this, letting the transformer learn by making mistakes. The transformer demonstrated its scalability by reaching final energies within 2% of the genuine ground-state energy, even when the system size rose to eight spins.
The Future of Quantum Control
The field of quantum control has advanced significantly with the success of the quantum transformer. In both Markovian and non-Markovian situations, this architecture offers a reliable, scalable, and fast feedback solution, which creates new opportunities for stabilizing complicated quantum states. These attention-based models’ capacity to manage extended context windows and a variety of noise profiles might be the key to realistic quantum error correction and state preparation as quantum hardware develops further. These models are anticipated to be tested experimentally in future studies, expanding the realm of what is feasible with the upcoming generation of quantum technology.
