Smart process analytics for predictive modeling

The state of technological advancement to address challenges in the manufacture of rAAV gene therapies

Current processes for the production of recombinant adeno-associated virus (rAAV) are inadequate to meet the surging demand for rAAV-based gene therapies. This article reviews recent advances that hold the potential to address current limitations in rAAV manufacturing. A multidisciplinary perspective on technological progress in rAAV production is presented, underscoring the necessity to move beyond incremental refinements and adopt a holistic strategy to address existing challenges. Since several recent reviews have thoroughly covered advancements in upstream technology, this article provides only a concise overview of these developments before moving to pivotal areas of rAAV manufacturing not well covered in other reviews, including analytical technologies for rapid and high-throughput measurement of rAAV quality attributes, mathematical modeling for platform and process optimization, and downstream approaches to maximize efficiency and rAAV yield. Novel technologies that have the potential to address the current gaps in rAAV manufacturing are highlighted. Implementation challenges and future research directions are critically discussed.

Design space determination of pharmaceutical processes: Effects of control strategies and uncertainty

The identification of process Design Space (DS) is of high interest in highly regulated industrial sectors, such as pharmaceutical industry, where assurance of manufacturability and product quality is key for process development and decision-making. If the process can be controlled by a set of manipulated variables, the DS can be expanded in comparison to an open-loop scenario, where there are no controls in place. Determining the benefits of control strategies may be challenging, particularly when the available model is complex and computationally expensive - which is typically the case of pharmaceutical manufacturing. In this study, we exploit surrogate-based feasibility analysis to determine whether the process satisfies all process constraints by manipulating the process inputs and reduce the effect of uncertainty. The proposed approach is successfully tested on two simulated pharmaceutical case studies of increasing complexity, i.e., considering (i) a single pharmaceutical unit operation, and (ii) a pharmaceutical manufacturing line comprised of a sequence of connected unit operations. Results demonstrate that different control actions can be effectively exploited to operate the process in a wider range of inputs and mitigate uncertainty.

Learning heterogeneous reaction kinetics from X-ray videos pixel by pixel

Reaction rates at spatially heterogeneous, unstable interfaces are notoriously difficult to quantify, yet are essential in engineering many chemical systems, such as batteries1 and electrocatalysts2. Experimental characterizations of such materials by operando microscopy produce rich image datasets3-6, but data-driven methods to learn physics from these images are still lacking because of the complex coupling of reaction kinetics, surface chemistry and phase separation7. Here we show that heterogeneous reaction kinetics can be learned from in situ scanning transmission X-ray microscopy (STXM) images of carbon-coated lithium iron phosphate (LFP) nanoparticles. Combining a large dataset of STXM images with a thermodynamically consistent electrochemical phase-field model, partial differential equation (PDE)-constrained optimization and uncertainty quantification, we extract the free-energy landscape and reaction kinetics and verify their consistency with theoretical models. We also simultaneously learn the spatial heterogeneity of the reaction rate, which closely matches the carbon-coating thickness profiles obtained through Auger electron microscopy (AEM). Across 180,000 image pixels, the mean discrepancy with the learned model is remarkably small (<7%) and comparable with experimental noise. Our results open the possibility of learning nonequilibrium material properties beyond the reach of traditional experimental methods and offer a new non-destructive technique for characterizing and optimizing heterogeneous reactive surfaces.

Bayesian Symbolic Learning to Build Analytical Correlations from Rigorous Process Simulations: Application to CO2 Capture Technologies

Process modeling has become a fundamental tool to guide experimental work. Unfortunately, process models based on first principles can be expensive to develop and evaluate, and hard to use, particularly when convergence issues arise. This work proves that Bayesian symbolic learning can be applied to derive simple closed-form expressions from rigorous process simulations, streamlining the process modeling task and making process models more accessible to experimental groups. Compared to conventional surrogate models, our approach provides analytical expressions that are easier to communicate and manipulate algebraically to get insights into the process. We apply this method to synthetic data obtained from two basic CO₂ capture processes simulated in Aspen HYSYS, identifying accurate simplified interpretable equations for key variables dictating the process economic and environmental performance. We then use these expressions to analyze the process variables' elasticities and benchmark an emerging CO₂ capture process against the business as usual technology.

Soft-Sensor for Class Prediction of the Percentage of Pentanes in Butane at a Debutanizer Column

Refineries are complex industrial systems that transform crude oil into more valuable subproducts. Due to the advances in sensors, easily measurable variables are continuously monitored and several data-driven soft-sensors are proposed to control the distillation process and the quality of the resultant subproducts. However, data preprocessing and soft-sensor modelling are still complex and time-consuming tasks that are expected to be automatised in the context of Industry 4.0. Although recently several automated learning (autoML) approaches have been proposed, these rely on model configuration and hyper-parameters optimisation. This paper advances the state-of-the-art by proposing an autoML approach that selects, among different normalisation and feature weighting preprocessing techniques and various well-known Machine Learning (ML) algorithms, the best configuration to create a reliable soft-sensor for the problem at hand. As proven in this research, each normalisation method transforms a given dataset differently, which ultimately affects the ML algorithm performance. The presented autoML approach considers the features preprocessing importance, including it, and the algorithm selection and configuration, as a fundamental stage of the methodology. The proposed autoML approach is applied to real data from a refinery in the Basque Country to create a soft-sensor in order to complement the operators' decision-making that, based on the operational variables of a distillation process, detects 400 min in advance with 98.925% precision if the resultant product does not reach the quality standards.