Quantum enhanced ensemble GANs for anomaly detection in continuous biomanufacturing

TL;DR

This work presents a novel framework for unsupervised anomaly detection in continuous biomanufacturing based on an ensemble of generative adversarial networks (GANs), and evaluates the impact of using a hybrid quantum/classical GAN approach with both a simulated quantum circuit and a real photonic quantum processor on anomaly detection performance.

Abstract

The development of continuous biomanufacturing processes requires robust and early anomaly detection, since even minor deviations can compromise yield and stability, leading to disruptions in scheduling, reduced weekly production, and diminished economic performance. These processes are inherently complex and exhibit non-linear dynamics with intricate relationships between process variables, thus making advanced methods for anomaly detection essential for efficient operation. In this work, we present a novel framework for unsupervised anomaly detection in continuous biomanufacturing based on an ensemble of generative adversarial networks (GANs). We first establish a benchmark dataset simulating both normal and anomalous operation regimes in a continuous process for the production of a small molecule. We then demonstrate the effectiveness of our GAN-based framework in detecting anomalies caused by sudden feedstock variability. Finally, we evaluate the impact of using a hybrid quantum/classical GAN approach with both a simulated quantum circuit and a real photonic quantum processor on anomaly detection performance. We find that the hybrid approach yields improved anomaly detection rates. Our work shows the potential of hybrid quantum/classical approaches for solving real-world problems in complex continuous biomanufacturing processes.

Figures (9)

• Figure 1: The anomaly detection workflow using an ensemble of GANs proposed in this study. A: first, using the KTB1 model for simulating the production process for lovastatin, a pharmaceutical ingredient, we produce a set of time series representing both normal and anomalous operating regimes. B: we train an ensemble of generative adversarial networks (GANs), each consisting of a generator ($G$) and discriminator ($D$) neural network, on the data from the normal regime. The generator is trained to map samples from an input probability distribution to synthetic data. Our work compares GANs trained with three different types of distributions: a classical Gaussian distribution, a simulated quantum distribution, and a real quantum distribution produced by an ORCA Computing PT-2 system. C: During the training process, each discriminator learns a decision boundary distinguishing between the normal operation data and other data. An ensemble of discriminators can thus be used as a detection anomaly module providing a set of learned criteria for flagging anomalies. If any discriminator in the ensemble flags a time series as anomalous, then the anomaly detection module labels the time series as an anomaly.
• Figure 2: Process flow diagram of the KTB-1 simulation model. The upstream section focuses on enhancing Lovastatin production (with V for valves, R for reactors, P for pumps, and HC for hydrocyclones), while the downstream section is dedicated to Lovastatin purification and includes the associated control system for dynamic process management (where C stands for centrifuge, T for tank, and NF for nanofiltration).
• Figure 3: Our anomaly detection scheme uses generative adversarial networks, in which two neural networks are trained in an adversarial way. The generator is trained to map samples from an initial distribution to realistic synthetic data. In this work, we investigate using either a classical Gaussian distribution or a quantum distribution. The discriminator learns to distinguish between the real data and the synthetic data. After training, we discard the generator and use the discriminator to determine whether an input time series is anomalous.
• Figure 4: Our experiments using a quantum processor use photonic quantum processors implemented with a time-bin architecture. As shown in a, in these architectures single photons are sequentially sent into a network of optical delay lines with programmable coupling coefficients. This creates an entangled state between different time bins which can measured by a photon number resolving detector. The two-loop architecture shown in a) implements the quantum optical circuit shown in b), where each gate is a programmable beam splitter between two optical modes.
• Figure 5: Receiver operating characteristic (ROC) curves of our anomaly detection algorithm based on an ensemble of 30 GAN discriminators. We consider GANs that use a classical Gaussian latent space or a quantum distribution produced by the quantum interference between identical photons. The quantum distribution is generated either by a real or by a simulated quantum processor. The GANs differ only in the process that produced the latent vectors sent to the generator. The solid line and shaded areas correspond respectively to the mean and standard deviation calculated by sampling 30 models 20 times from an ensemble of 120 trained models.