Milad Ramezankhani

Despite the successful implementations of physics-informed neural networks in different scientific domains, it has been shown that for complex nonlinear systems, achieving an accurate model requires extensive hyperparameter tuning, network architecture design, and costly and exhaustive training processes. To avoid such obstacles and make the training of physics-informed models less precarious, in this paper, a data-driven multi-fidelity physics-informed framework is proposed based on transfer learning principles. The framework incorporates the knowledge from low-fidelity auxiliary systems and limited labeled data from target actual system to significantly improve the performance of conventional physics-informed models. While minimizing the efforts of designing a complex task-specific network for the problem at hand, the proposed settings guide the physics-informed model towards a fast and efficient convergence to a global optimum. An adaptive weighting method is utilized to further enhance the optimization of the model composite loss function during the training process. A data-driven strategy is also introduced for maintaining high performance in subdomains with significant divergence between low- and high-fidelity behaviours. The heat transfer of composite materials undergoing a cure cycle is investigated as a case study to demonstrate the proposed framework’s performance compared to conventional physics-informed models.

Unprecedented advances in Machine Learning (ML), cloud computing and sensory technology promise to enable the manufacturing industry to respond rapidly to changes in marked needs while maintaining product quality and minimizing costs. Despite the unparalleled advantages that ML offers, critical limiting factors have prevented the exhaustive cultivation of ML in advanced manufacturing. Constant shifts in the process configuration and lack of sufficient fully-descriptive data restrict the performance of predictive ML models. This paper proposes to partly address these shortfalls with an active transfer learning (ATL) model that is applied to an aerospace composites manufacturing case study. The proposed ATL framework requires 1) developing an AI-based optimal experimental design using Active Learning (AL) to maximize the information gain from the limited number of allowable manufacturing trials, and 2) equipping the manufacturing process with a robust Transfer Learning (TL) model that is trained on limited available data and is immune to shifts in the process settings. The results suggest that uncertainty-based AL approaches can significantly decrease the dependency on large datasets for obtaining accurate process models. Furthermore, in comparison with traditional TL approaches, the proposed framework represents a practical solution to further reduce the necessity for generating expensive data in advanced manufacturing applications for developing reliable and transferable predictive models.

The integration of advanced manufacturing processes with ground-breaking Artificial Intelligence methods continue to provide unprecedented opportunities towards modern cyber-physical manufacturing processes, known as smart manufacturing or Industry 4.0. However, the “smartness” level of such approaches closely depends on the degree to which the implemented predictive models can handle uncertainties and production data shifts in the factory over time. In the case of change in a manufacturing process configuration with no sufficient new data, conventional Machine Learning (ML) models often tend to perform poorly. In this article, a transfer learning (TL) framework is proposed to tackle the aforementioned issue in modeling smart manufacturing. Namely, the proposed TL framework is able to adapt to probable shifts in the production process design and deliver accurate predictions without the need to re-train the model. Armed with sequential unfreezing and early stopping methods, the model demonstrated the ability to avoid catastrophic forgetting in the presence of severely limited data. Through the exemplified industry-focused case study on autoclave composite processing, the model yielded a drastic (88%) improvement in the generalization accuracy compared to the conventional learning, while reducing the computational and temporal cost by 56%.