Deep learning-based surrogate model for three-dimensional patient-specific computational fluid dynamics

Abstract

Optimization and uncertainty quantification have been playing an increasingly important role in computational hemodynamics. However, existing methods based on principled modeling and classic numerical techniques have faced significant challenges, particularly when it comes to complex 3-dimensional (3-D) patient-specific shapes in the real world. First, it is notoriously challenging to parameterize the input space of arbitrarily complex 3-D geometries. Second, the process often involves massive forward simulations, which are extremely computationally demanding or even infeasible. We propose a novel deep learning surrogate modeling solution to address these challenges and enable rapid hemodynamic predictions. Specifically, a statistical generative model for 3-D patient-specific shapes is developed based on a small set of baseline patient-specific geometries. An unsupervised shape correspondence solution is used to enable geometric morphing and scalable shape synthesis statistically. Moreover, a simulation routine is developed for automatic data generation by automatic meshing, boundary setting, simulation, and post-processing. An efficient supervised learning solution is proposed to map the geometric inputs to the hemodynamics predictions in latent spaces. Numerical studies on aortic flows are conducted to demonstrate the effectiveness and merit of the proposed techniques.

Deep Learning to Reveal the Distribution and Diffusion of Water Molecules in Fuel Cell Catalyst Layers

“…For more control variables or conditions (e.g., different temperature, different wettability, and different surface roughness of carbon), the existing deep learning framework required recalculation and the support of new datasets, which was the biggest limitation of the proposed learning framework. Based on this limitation, combined with recent advances in the field of deep learning, , we are developing unsupervised learning network frameworks that automatically generate data for different conditions. In the subsequent work, the unsupervised learning framework will automatically generate point cloud samples of water molecule transport for different cases such as the surface roughness of the carbon substrate based on few MD simulation arithmetics.…”

Differentiable hybrid neural modeling for fluid-structure interaction

“…The use of deep neural networks (DNNs) for modeling complex nonlinear dynamics of fluid or FSI systems has been actively explored in recent years. By directly learning the input-output relationships in either high-dimensional or reduced-order space using data-driven techniques such as proper orthogonal decomposition (POD) and convolutional autoencoder methods, mapping functions have been identified for various scenarios, including flow past cylinders [9,10], airfoil aerodynamics [11,12], biological flows [13,14], and RANS/LES closure problems [15][16][17][18], among others. To handle irregular domains with unstructured data, graph neural networks (GNNs) have been used to build fast neural simulators for fluid and FSI dynamics [19,20].…”

FLUID-GPT (Fast Learning to Understand and Investigate Dynamics with a Generative Pre-Trained Transformer): Efficient Predictions of Particle Trajectories and Erosion

“…In recent years, there has been a rapid paradigm shift from “traditional” computational approaches to data-driven science, largely due to advances in computational power and the increasing availability of data. In the context of CFD and erosion calculations, recent machine learning (ML) approaches have shown remarkable accuracy and efficiency in a variety of systems. − In particular, a previous work by us utilized a hybrid long- and short-term memory (LSTM) with a three-dimensional convolutional neural network (3D CNN) to predict particle trajectories and surface erosion, respectfully, in an industrial-scale boiler header . While our approach was able to successfully predict surface erosion using only five initial conditions as input, the LSTM training was computationally expensive due to its recurrence-based architecture, taking approximately 26 h on 32 parallel CPUs.…”