Foundation Model Achieves Unified Characterization of Complex Optical Quantum States

Characterizing the properties of quantum states of light is crucial for advancing quantum technologies, yet current methods struggle to provide a comprehensive understanding of complex states without demanding extensive measurement. Xiaoting Gao from Peking University, Yan Zhu from The University of Hong Kong, Feng-Xiao Sun from Beijing University of Posts and Telecommunications, and colleagues now present the first foundation model capable of unified characterization of optical quantum states across a wide range of complexities. This model overcomes limitations of previous approaches by accurately predicting key properties, such as fidelity and Wigner negativity, for experimentally relevant states including those exhibiting strong non-Gaussian behaviour, multiple modes, and high levels of squeezing. By enabling efficient characterization from limited data, this work establishes a unified framework that promises to accelerate progress in optical quantum information computation, communication, and precision measurement.

Scientists have now developed a foundation model capable of unified characterization of optical quantum states across a wide range of complexities. This model overcomes limitations of previous approaches by accurately predicting key properties, such as fidelity and Wigner negativity, for experimentally relevant states including those exhibiting strong non-Gaussian behaviour, multiple modes, and high levels of squeezing. By enabling efficient characterization from limited data, this work establishes a unified framework that promises to accelerate progress in optical quantum information computation, communication, and precision measurement.

Predicting Quantum States Via Foundation Models

Scientists developed a foundation model for characterizing optical quantum states, addressing the challenge of predicting properties for complex, multimode non-Gaussian systems without requiring full quantum state tomography. The study pioneers a pretraining and fine-tuning strategy, beginning with the prediction of homodyne measurement statistics from computationally tractable optical states. This initial stage establishes self-supervised representations of lower complexity states, forming the basis for generalization to more challenging regimes. The team then evaluated the model’s ability to predict properties of increasingly complex states without additional training, assessing its out-of-distribution generalization capabilities. This rigorous testing phase demonstrated the potential for accurate prediction even with limited training data.

Foundation Model Characterizes Complex Quantum States

Scientists have developed a foundation model capable of characterizing optical quantum states across a wide range of complexities, defined by non-Gaussianity, the number of modes, and the degree of squeezing. The work establishes a unified framework for predicting diverse quantum properties within a single neural network architecture, employing a pretraining and fine-tuning strategy. The team constructed optical quantum states using a precise method for controlling non-Gaussian properties. Experiments demonstrate the model’s ability to accurately predict key quantum properties, such as Wigner negativity and quantum fidelity, after fine-tuning on limited data. This framework establishes a pathway toward scalable, generalizable machine learning models for optical quantum state characterization, mirroring the success of foundation models in other scientific domains.

Optical State Characterization With Machine Learning

This research introduces a foundational model for characterizing optical states across a broad spectrum of complexity, defined by non-Gaussianity, the number of modes, and the degree of squeezing. The team demonstrates that a single model, initially trained on simpler states, can accurately characterize more complex states without requiring complete state tomography, a traditionally demanding process. The model achieves this by learning a compact representation of quantum states from limited measurement data, effectively predicting properties like fidelity and Wigner negativity. Importantly, the learned representation organizes non-Gaussian states according to established theoretical predictions. This work establishes a unified framework for efficient state characterization, which is particularly valuable in scenarios where obtaining large labelled datasets is challenging.