The 22nd ACM conference on Knowledge Discovery and Data Mining (KDD) is a main venue for academic and industry research in data mining, machine learning, information retrieval, knowledge discovery, large-scale data analytics and big data. KDD 2016 will take place next week, August 15-17, in San Francisco, California.
This year, Autodesk researchers will present the paper, Convolutional Neural Networks for Steady Flow Approximation, co-authored by Wei Li, Francesco Iorio and former Autodesk intern Xiaoxiao Guo at the conference in the Applied Data Science Track. This paper uses convolutional neural networks to build fast CFD surrogate models for interactive design and design space exploration.
In aerodynamics and fluid dynamics related analysis problems, flow fields are traditionally simulated using computational fluid dynamics (CFD) solvers. However, CFD simulation is usually a computationally expensive, memory demanding and time consuming iterative process. These drawbacks of CFD limit opportunities for design space exploration and forbid interactive design.
Inspired by recent progress of deep learning, this work proposed a general and flexible approximation model for real-time prediction of non-uniform steady laminar flow in a 2D or 3D domain based on convolutional neural networks (CNNs).
This paper explored alternatives for the geometry representation and the network architecture of CNNs. It shows that convolutional autoencoder can estimate the velocity field two orders of magnitude faster than a GPU-accelerated CFD solver and four orders of magnitude faster than a CPU-based CFD solver at a cost of a low error rate.
The encoding part of autoencoder takes the geometry input, and stacked convolution layers extract geometry features. The 'inverse' operation of convolution, called deconvolution, is then used to construct multiple stacked decoding layers. Deconvolution layers multiply each input value by a filter elementwise, and sum over the resulting output windows. In other words, a deconvolution layer is a convolution layer with the forward and backward passes reversed.
This approach can provide immediate feedback for real-time design iterations at the early stage of design. Compared with existing approximation models in the aerodynamics domain, CNNs enable an efficient estimation for the entire velocity field. Furthermore, designers and engineers can directly apply the CNN approximation model in their design space exploration algorithms without training extra lower-dimensional surrogate models.