In this study, we propose a population-based, data-driven intelligent controller that leverages neural-network-based digital twins for hypothesis testing. Initially, a diverse set of control laws is generated using genetic programming with the digital twin of the system, facilitating a robust response to unknown disturbances. During inference, the trained digital twin is utilized to virtually test alternative control actions for a multi-objective optimization task associated with each control action. Subsequently, the best policy is applied to the system. To evaluate the proposed model predictive control pipeline, experiments are conducted on a multi-mode heat transfer test rig. The objective is to achieve homogeneous cooling over the surface, minimizing the occurrence of hot spots and energy consumption. The measured variable vector comprises high dimensional infrared camera measurements arranged as a sequence (655,360 inputs), while the control variable includes power settings for fans responsible for convective cooling (3 outputs). Disturbances are induced by randomly altering the local heat loads. The findings reveal that by utilizing an evolutionary algorithm on measured data, a population of control laws can be effectively learned in the virtual space. This empowers the system to deliver robust performance. Significantly, the digital twin-assisted, population-based model predictive control (MPC) pipeline emerges as a superior approach compared to individual control models, especially when facing sudden and random changes in local heat loads. Leveraging the digital twin to virtually test alternative control policies leads to substantial improvements in the controller’s performance, even with limited training data.

Objective of this study is to investigate the effect of secondary motion of particles in multiphase gas-solid flows parametrically and test the relative impacts of particle shape and orientation information on particle distribution. For that purpose, predictive accuracies of simplified drag coefficient models are assessed for the conditions relevant to a wood recovery plant operating at dilute flow regime. After demonstrating the strong impact of the shape and orientation information on the force balance for single particles, we compared the steady state Eulerian-Lagrangian simulation results for particle volume fractions, residence times and particle diameter distributions within the chamber for different (i) superficial gas velocities (5 m/s, 7.5 m/s), (ii) orientation tendencies and (iii) particle shapes. Transient simulations are performed until the system reaches steady state conditions by monitoring the mass flow rates of the particulate phases leaving the chamber. The secondary motion of non-spherical particles is represented by stochastic sampling from the available experimental data. Analysis of the force balance on single particles revealed log-scale variations if the orientation of the particles with respect to flow fluctuates. Variations in the single particle force balances are found to be still visible in the CFD analysis, where the secondary motion of particles drastically changed the particle distribution in the chamber. The native non-spherical model which only accounts for the shape correction was found to over-predict the entrainment, leading to a significantly different particle volume fraction and diameter distributions. Spherical particle assumption also caused significant errors in the particle distribution, which increases as aspect ratio of the cylindrical particle diverges from one. Results show that particle orientation statistics are extremely important to capture the particle mixing and segregation patterns at dilute regime, which cannot be captured with such simplifying assumptions.

The inherent dependency of deep learning models on labeled data is a well-known problem and one of the barriers that slows down the integration of such methods into different fields of applied sciences and engineering, in which experimental and numerical methods can easily generate a colossal amount of unlabeled data. This paper proposes an unsupervised domain adaptation methodology that mimics the peer review process to label new observations in a different domain from the training set. The approach evaluates the validity of a hypothesis using domain knowledge acquired from the training set through a similarity analysis, exploring the projected feature space to examine the class centroid shifts. The methodology is tested on a binary classification problem, where synthetic images of cubes and cylinders in different orientations are generated. The methodology improves the accuracy of the object classifier from 60% to around 90% in the case of a domain shift in physical feature space without human labeling.

The inherent dependency of deep learning models to labeled data is a well-known problem and one of the barriers that slows down the integration of such methods into different fields of applied sciences and engineering, in which experimental and numerical methods can easily generate a colossal amount of unlabeled data. This paper proposes an unsupervised domain adaptation methodology that mimics the peer review process to label new observations in a different domain from the training set. The approach evaluates the validity of a hypothesis using domain knowledge acquired from the training set through a similarity analysis, exploring the projected feature space to examine the class centroid shifts. The methodology is tested on a binary classification problem, where synthetic images of cubes and cylinders in different orientations are generated. The methodology improves the accuracy of the object classifier from 60% to around 90% in the case of a domain shift in physical feature space without human labeling.