Toward Transparent and Controllable Quantum Generative Models Authors: Jinkai Tian, Wenjing Yang Published: 2024 ABSTRACT Quantum generative models have shown promise in fields such as quantum chemistry, materials science, and optimization. However, their practical utility is hindered by a significant challenge: the lack of interpretability. In this work, we introduce model inversion to enhance both the interpretability and controllability of quantum generative models. Model inversion allows for tracing generated quantum states back to their latent variables, revealing the relationship between input parameters and generated outputs. We apply this method to models generating ground states for Hamiltonians, such as the transverse-field Ising model (TFIM) and generalized cluster Hamiltonians, achieving interpretability control without retraining the model. Experimental results demonstrate that our approach can accurately guide the generated quantum states across different quantum phases. This framework bridges the gap between theoretical models and practical applications by providing transparency and fine-tuning capabilities, particularly in high-stakes environments like drug discovery and material design. 1. Introduction Quantum generative models [1] have shown significant potential in fields such as drug discovery and materials science [2], leveraging quantum mechanics to efficiently generate complex quantum states or data samples [3]. However, these models face unique interpretability challenges due to quantum properties like superposition and entanglement, which contribute to high entropy in the state space and complicate output interpretation. A primary challenge is the difficulty in interpreting the relationship between input variables (e.g., Hamiltonian parameters) and output quantum states [4,5,6,7,8]. Unlike classical models, where layers and parameters can often be directly analyzed, quantum models operate within highly abstract quantum state spaces. The “black box” nature of quantum models limits our understanding of their internal operations, making it challenging to control or verify specific outcomes, especially in high-stakes applications. This lack of interpretability has significant practical implications. In critical fields like drug discovery and materials science, understanding how quantum states are generated is vital for ensuring trustworthy results that meet stringent requirements. Without the ability to interpret the relationships between input parameters and generated quantum states, controlling these models to achieve desired outcomes becomes extremely difficult, undermining their reliability and broader adoption. While classical generative models also face interpretability challenges, various techniques—such as visualization tools, simplified architectures, and model inversion—have been developed to enhance understanding [9,10]. However, these methods are often insufficient or inapplicable to quantum models due to their probabilistic nature and complex non-linearities, creating a critical gap in our ability to understand and control quantum generative models. To address these challenges, we propose applying model inversion techniques to quantum generative models. Model inversion maps generate quantum states back to their latent variables, revealing the underlying relationships between inputs and outputs. This approach allows users to trace how specific quantum states are produced, offering a mechanism to control and fine-tune outputs by adjusting latent variables. Moreover, it enables the identification of specific directions in the latent space that correlate with distinct features in the generated samples, allowing for targeted interventions in the generative process. By enhancing both interpretability and control, model inversion directly addresses the critical gap in current research. It provides a way to understand and manipulate the complex, probabilistic nature of quantum models, building trust and enabling more precise control over outputs—being especially important in applications where fine-tuning quantum states is essential. In this study, we apply model inversion techniques to quantum generative models, focusing on generating ground states for Hamiltonians such as the transverse-field Ising model (TFIM) and generalized cluster Hamiltonians. Our approach not only enhances interpretability but also provides a pathway for controlling outputs, making quantum models more reliable and adaptable for practical applications. By addressing interpretability and control challenges, this work contributes to bridging the gap between the theoretical promise of quantum generative models and their practical utility. With model inversion, we offer a method that allows researchers and practitioners to explore the internal workings of quantum models more transparently, providing both control and flexibility—critical elements for deploying quantum generative models in real-world applications, where precision, reliability, and interpretability are paramount. 2. Preliminary 2.1. Quantum Generative Models Quantum generative models have emerged as powerful tools for simulating and generating quantum states across various quantum systems. These models leverage quantum computing’s inherent ability to handle complex, high-dimensional state spaces, offering a range of approaches for state generation. One such class is the Quantum Circuit Born Machine (QCBM) [11,12], which uses quantum circuits to learn and represent probabilistic distributions of quantum states. QCBMs are particularly effective at learning quantum data distributions and can be applied to quantum state preparation and quantum machine learning tasks. Another prominent class is Quantum Generative Adversarial Networks (QGANs) [13,14], which apply a competitive learning process similar to classical GANs, where a quantum generator creates states that a discriminator (classical or quantum) evaluates, allowing for high-fidelity state generation. Each of these quantum generative models has demonstrated success in tasks such as quantum state preparation, quantum state tomography, and quantum simulations. However, despite these advancements, significant challenges remain. In particular, understanding and controlling the generated quantum states pose serious difficulties due to the complexity of quantum systems and the probabilistic nature of quantum mechanics. In this context, studying specific quantum systems, such as the transverse-field Ising model (TFIM) and generalized cluster Hamiltonian, provides a structured approach to evaluating and understanding quantum generative models, as they represent well-known quantum phase transition systems that can serve as benchmarks for state generation tasks. 2.2. Quantum Phase Transitions Quantum phase transitions are phase transitions driven by quantum fluctuations, which occur at absolute zero (0 Kelvin) as a parameter in the Hamiltonian changes [15]. Two prominent models in the study of quantum phase transitions are the transverse-field Ising model (TFIM) and the generalized cluster Hamiltonian. The Hamiltonian for the one-dimensional TFIM is given by (1)H=−J∑iσziσzi+1−h∑iσxi, where J represents the interaction strength between neighboring spins, h denotes the transverse magnetic field strength, and σxi and σzi are the Pauli-X and Pauli-Z operators acting on the i-th spin, respectively. In the TFIM, quantum phase transitions occur at the critical point where the relative strength between the transverse magnetic field h and the spin–spin interaction J changes. Specifically, for the one-dimensional TFIM, the critical point is reached at J=h, leading to a quantum phase transition from an ordered phase (J>h) to a disordered phase (J