Abstract The vast majority of stochastic model updating research thus far has concentrated around Bayesian sampling approaches. While these existing approaches work well, their extension to the time-series is not clear and their dependency on sampling makes them computationally unfeasible for some models. Here we share our recent work which migrates away from Bayesian posterior inference to a machine learning approach which amortizes the posterior directly upon input-output training data for efficient inference with observations obtained online.

Abstract Certification is all about attesting to the reliability of a man-made product or system. It requires extensive testing under conditions that will be encountered during the operation of the asset, and even under some conditions that exceed normal operational scenarios. Because of this, certification campaigns can last for years, cost millions, result in the damage or loss of assets and sometimes be dangerous to the people involved. Certification by simulation has emerged as a concept for alleviating the burden of conducting the entire certification campaign physically. However, regulators are still unwilling to trust computational models alone for pronouncements about system safety, reliability and performance. The regulations themselves are unclear about the evidence upon which a simulation campaign should be classed as acceptable to substitute experiments.

At their advent digital twins promised unprecedented level of reliability in the representation of the physical asset of interest, seemingly making them the ideal candidate to carry the burden of certification by simulation. Although reliability is universally accepted as a key characteristic of digital twins, there is little consensus on how it should be achieved in practice.

This talk will explore the issue of reliability of digital twins, whether digital twins are the long-sought-for solution to the problem of certification by simulation and what is still missing in the field.

Abstract Smart manufacturing systems require adaptive workload allocation and operating mode selection due to dynamic production demands, leading to varying device degradation. Load balancing is critical for joint operation and maintenance (O&M) planning, yet predictive control for multi-mode systems remains understudied. This paper proposes a digital twin-driven proactive O&M framework, integrating task and health data to dynamically regulate system health. A multi-stage integer linear programming (MSILP) model with generalized disjunctive programming (GDP) constraints optimizes operating modes and maintenance planning, minimizing costs while ensuring reliability by considering component dependencies, resources, and degradation. A matheuristics algorithm enables rolling horizon control for customized demands. A subsea production system case study demonstrates how task requirements and rolling horizons impact performance in customer-oriented markets.

Abstract Likelihood-free inference methods have been widely adopted, but they face significant challenges in updating multi-level computational models that have hierarchically embedded sub-models. This difficulty arises from the lack of direct observations of the quantities of interest of the sub-models. In addition, recent advancements in sensing and image processing technologies allow for the collection of a substantial amount of video monitoring data through non-contact sensing techniques. The implicit and very complicated relationship between the uncertain model parameters and video monitoring data adds an additional layer of challenge to the updating of multi-level models. This research overcomes these challenges by proposing an innovative Recursive Inference method based on Invertible Neural Networks (RINN) for multi-level models. The proposed RINN framework first compresses the high-dimensional video monitoring data into low-dimension latent-space data using a convolutional autoencoder. Using synthetic video monitoring data generated by considering various uncertainty sources in the computational simulation models, a likelihood-free inference model is then trained through a conditional invertible neural network (cINN) and a summary neural network. This model efficiently approximates the posterior distributions of uncertain model parameters without evaluating the computationally intractable likelihood function, given any latent-space data obtained from the autoencoder. To facilitate continuous monitoring and model updating over an extended monitoring period, this study further proposes a recursive model updating strategy that integrates the cINN-based likelihood-free inference with the particle filtering method. The updating of a degradation model of a miter gate application is employed as an example throughout this paper to explain and demonstrate the efficacy of the proposed RINN framework. The results of the case study show that the RINN is able to effectively reduce uncertainty in the degradation model parameters through strain video monitoring data of a miter gate, and thereby increase the confidence in the remaining useful life (RUL) estimation using the updated degradation model. 

Abstract Digital Twins (DT) are critical for guiding high-consequence decision making on complex engineering systems and critical infrastructure and for understanding and predicting their behavior in operation. Physics-Enhanced Machine Learning (PEML) strategies offer a powerful tool for enabling the development of Robust and Interpretable DT based on the integration of physics-based models, expert and domain knowledge, and real-world data. This integration is a nontrivial task since: (i) real-world data is often expensive to acquire and difficult to measure, small in volume, heterogeneous, gappy, noisy, multimodal, and multi-fidelity, with different spatial and temporal resolutions; (ii) physics-based models can be multi-fidelity, multi-scale, high-dimensional, coupled, deterministic or stochastic, often compromised by errors. During this lecture, PEML will be introduced and three broad groups of PEML approaches will be discussed: physics-guided, physics-encoded and physics-informed. Recent strategies developed for Digital Twinning for Reliability Engineering across a wide range of real-world applications will be presented. Open challenges and opportunities will be discussed. A short intro to PEML can be found here: https://iopscience.iop.org/article/10.1088/1742-6596/2909/1/012034/meta

Abstract To fulfill the high-temperature, high-voltage, and high-frequency applications of Wide Bandgap Semiconductor (WBG) power electronics, an excellent interconnect technology that can withstand harsh environments for a long time are necessary in WBG packaging. With its benefits of low processing temperatures, exceptional electro-thermo-mechanical performance, and high process flexibility, nano/micro metal particle sintering technology is gaining increased attention, particularly in automotive and power electronics packaging applications. In this talk, the digital twin assisted nano/micro-Ag and Cu particle and paste material preparations and sintering process, property characterization, reliability testing and applications will be introduced, respectively. Finally, the challenges and outlook of promising DT assisted sintering technology will be summarized. 

Abstract Estimation of multimodal and sharp posteriors with nonlinear dependencies as well as the associated model evidence remains a critical challenge in many Bayesian model inference tasks such as model parameter calibration, model selection and model averaging. Bayesian Quadrature (BQ) based on approximating the logarithm of likelihood with a Gaussian Process surrogate model has been proven to be a promising scheme for multimodal inference, but the mechanism behind it has not yet been sufficiently investigated. This lecture presents a substantially improved BQ scheme for addressing this challenge. By exploring the mechanism of exponential impact behind this, I first answer the questions ``why it works?'', as well as ``can it work better, and how?'' This mechanism then motivates the development of a simpler but more effective BQ method informed by the exponential impact. This BQ method is equipped with measures of prediction uncertainties and active learning, driven by two new acquisition functions, which have insightful interpretations, closed-form expressions and sound performance. Further, a transitional learning scheme based on adaptive tempering is developed and embedded into the developed BQ method, allowing for adaptive inference of sharp posteriors with desired accuracy. I also presents several benchmark and engineering case studies to demonstrate the high efficiency and robustness of the proposed method to Bayesian model inference problems with complex posteriors.