Search for: All records

Abstract Physics-Informed Neural Networks (PINNs) have opened new possibilities for solving partial differential equations (PDEs) by embedding physical laws directly into the learning process. However, despite their flexibility, traditional PINNs often struggle to capture sharp gradients and intricate solution features, which limits their effectiveness in many practical problems. In this work, we have introduced Gradient-Driven Physics-Informed Neural Networks (GDPINNs) that improve the ability of traditional PINNs to resolve sharp gradients. By incorporating gradient information directly into the loss function, GDPINNs better target regions where traditional PINNs typically fail. We validated the method on steady-state and transient heat conduction problems, including a central heating source and a sinusoidal boundary condition, and found strong agreement with reference solutions. To further understand the framework's capability, we applied it to a high-gradient steady-state and transient heat conduction problem, where GDPINNs show clear advantages over traditional PINNs and align closely with reference results. We also extended GDPINNs to incompressible laminar flow in a lid-driven cavity, demonstrating its broader applicability. In these cases, GDPINNs consistently provide higher accuracy and better capture critical solution features, highlighting their potential to improve PINNs-based approaches for complex physical problems with sharp gradients. 

The primary goal of sinonasal surgery is to improve a patient’s quality of life, which is generally achieved by enhancing drug delivery (eg, saline rinses, nasal steroids) and nasal airflow. Both drug delivery and nasal airflow are dependent on the anatomic structure of the sinonasal cavity and the relationship between this anatomy and airflow and drug delivery can be studied using computational fluid dynamics (CFD). CFD generally uses computed tomography scans and computational algorithms to predict airflow or drug delivery and can help us understand surgical outcomes and optimize drug delivery for patients. This study employs CFD to simulate nasal airflow dynamics and optimize drug delivery in the nasal cavity to highlight the utility of CFD for studying sinonasal disease. Utilizing COMSOL Multiphysics software, we developed detailed models to analyze changes in airflow characteristics before and after functional endoscopic sinus surgery, focusing on pressure distribution, velocity profiles, streamline patterns, and heat transfer. This research examines the impact of varying levels of nasal airway obstruction on airflow and heat transfer. In addition, we explore the characteristics of nasal drug delivery by simulating diverse spray parameters, including particle size, spray angle, and velocity. Our comprehensive approach allows for the visualization of drug particle trajectories and deposition patterns, providing crucial insights for enhancing surgical outcomes and improving targeted drug administration. By integrating patient-specific nasal cavity models and considering factors such as airway outlet pressure, this study offers valuable data on pressure cross-sections, flow rate variations, and particle behavior within the nasal passages. The findings of this research can be useful for both surgical planning and the development of more effective nasal drug delivery methods, potentially leading to enhanced clinical outcomes in respiratory treatment. 

Training a machine learning model with data following a meaningful order, i.e., from easy to hard, has been proven to be effective in accelerating the training process and achieving better model performance. The key enabling technique is curriculum learning (CL), which has seen great success and has been deployed in areas like image and text classification. Yet, how CL affects the privacy of machine learning is unclear. Given that CL changes the way a model memorizes the training data, its influence on data privacy needs to be thoroughly evaluated. To fill this knowledge gap, we perform the first study and leverage membership inference attack (MIA) and attribute inference attack (AIA) as two vectors to quantify the privacy leakage caused by CL. Our evaluation of 9 real-world datasets with attack methods (NN-based, metric-based, label-only MIA, and NN-based AIA) revealed new insights about CL. First, MIA becomes slightly more effective when CL is applied, but the impact is much more prominent to a subset of training samples ranked as difficult. Second, a model trained under CL is less vulnerable under AIA, compared to MIA. Third, the existing defense techniques like MemGuard and MixupMMD are not effective under CL. Finally, based on our insights into CL, we propose a new MIA, termed Diff-Cali, which exploits the difficulty scores for result calibration and is demonstrated to be effective against all CL methods and the normal training method. With this study, we hope to draw the community's attention to the unintended privacy risks of emerging machine-learning techniques and develop new attack benchmarks and defense solutions.