Scientific machine learning enables potential to design at ‘near-interactive speeds’
Researchers are developing deep learning methods to dramatically reduce the cost and turnaround of conceptual design computations for complex energy systems
Credit: Oden Institute
Researchers are developing deep learning methods to dramatically reduce the cost and turnaround of conceptual design computations for complex energy systems. The project, led by researchers from the Oden Institute for Computational Engineering and Sciences at the University of Texas at Austin, will demonstrate these methods for optimal shape design of aircraft wings. However, it is being developed as a platform that can be applied to improve design processes across a broad range of energy systems.
“The goal is to be able to execute design studies at near-interactive speeds,” says project lead Dr. Omar Ghattas of the Oden Institute and Professor of Mechanical Engineering and Geological Sciences at UT Austin. “This is in contrast to current state-of-the-art, where such design studies take weeks or months even with modern high performance computing power.”
The team of Omar Ghattas and Karen Willcox from the Oden Institute has been awarded a $1.6 million grant to develop these methods as part of the Department of Energy’s ARPA-E DIFFERENTIATE initiative, a new program seeking to “enhance the pace of energy innovation by incorporating machine learning into the energy technology development process.”
Alongside Joaquim Martins from Aerospace Engineering at the University of Michigan, the team are using ‘scientific machine learning’ – an approach that blends scientific computing with machine learning. They will apply their formidable computational modeling and simulation expertise to create efficient, accurate, and scalable deep neural network (DNN) representations of solution of design optimization problems.
Specifically, they will focus on the problem of optimizing the shape of aircraft wings to improve aerodynamic performance and efficiency.
But the methodology developed will be broadly applicable to design and control problems common to various fluid dynamic design problems, including wind turbines, turbine engines, ship hulls, submersibles, and automotive bodies.
The DIFFERENTIATE program identifies mathematical optimization problems common to many design processes, such as inverse design: tackling a problem in reverse, working backward from the desired performance (effect) to compute the design (cause).
Inverse problems are some of the most important problems in science and mathematics because they tell us about features that we cannot directly observe.
DNNs designed for use in a traditional forward problem solution approach (calculating the effect given the cause) are predicated on the availability of massive amounts of training data. When it comes to building a DNN for the inverse problem, generating that training data could quickly become prohibitively expensive, since it entails computing a large number of optimal designs, each requiring many forward model solutions.
That’s where the team bring in their new mathematical ideas, which exploit the low-dimensional structure underlying the inverse design problem to reduce the training data required and guide the training process.
“Our guiding principle is this: the simplest machine learning model that explains the optimal design data is the best,” Ghattas stressed.
“Not just to make the technology more widely accessible but because in doing so we avoid overfitting the model to the data, which leads to the inability to predict unseen data.”
“It’s about making aerodynamic design optimization accessible to a much broader set of industrial users,” Ghattas said. “We know these technologies can enable faster, cheaper and more energy efficient design and production methods. So they should also be universally available.”