MIE Seminar: Saketh Sridhara, Assistant Professor, Mechanical and Industrial Engineering UMass Amherst, "ML-driven representations in engineering design optimization"

Content

Abstract

Gradient-based design optimization methods involving continuous variables (representing the size, shape, and/or topology of objects) have been successfully applied across various disciplines, including mechanical, aerospace, and civil engineering, to design optimized components, devices, and structures. The design of engineered systems and assemblies often involve selection from discrete tabular data contained in engineering catalogs and handbooks—such as materials, bearings, springs, and motors—that are not amenable to gradient-based optimization. Further, most commercially available physics-based solvers today function as “black-boxes” and lack differentiability, further prohibiting gradient-based optimization in complex multi-physics settings. 

This talk presents a new computational design framework that leverages recent advances in machine learning and differentiable programming. Neural networks, in particular, variational autoencoders, transform discrete catalog data to continuous, differentiable representations, making them compatible with gradient-based optimization. Additionally, a differentiable, fast finite element solver is developed that enables automatic gradient computation via backpropagation. This end-to-end differentiable framework allows for simultaneous optimization of both continuous and discrete variables, such as integrated geometry and material optimization, leading to enhanced part performance, minimal material usage and sustainable design choices. Further, neural network-based geometric representations can, under certain conditions, find better local minima in complex design optimization problems. This capability is demonstrated in the design of “metamaterials” —multi-scale structures with exceptional mechanical properties such as negative Poisson’s ratio. Future directions in multi-physics and multi-scale design are also discussed.

Bio

Saketh Sridhara is an assistant professor in the department of mechanical and industrial engineering at UMass Amherst. His research focuses on developing novel design optimization frameworks that integrate multiple physics, materials, and length scales, to design high- performance components and devices across various engineering disciplines. His work leverages tools from computational mechanics, CAD, optimization, machine learning and additive manufacturing to advance the state of the art in mechanical design. Prior to joining UMass, Saketh received his Ph.D. from University of Wisconsin-Madison in mechanical engineering, with a minor in computer science.