Research
Generative Linkages
Authors: Aditya Palaparthi, Alex Guerra, Ryan P. Adams
Institution: Princeton University
Date: In-Progress
Info: Developing a SOTA deep generative model for inversely designing linkage mechanisms given a path of motion with differentiable simulation capabilities. In this work, we also gain an algebraic understanding of the design space of linkage mechanisms, enabling our generative model to efficiently navigate towards valid, optimal linkage mechanisms.
Discovery of Structural Form with Generative Flow Networks
Authors: Alex Guerra1, Aditya Palaparthi1, Robert Hawkins2, Sarah-Jane Leslie1
Institutions: 1Princeton University, 2Stanford University
Date: In-Progress
Info: Human knowledge is richly structured. We know that dogs are more similar to other animals than they are to, say, mushrooms (hierarchical), that the sun rises and sets in roughly 24 hour cycles (periodic), that the numeral 5 is less than 9 (linear), and that our friends belong to different fami- lies (clustered). But where does this structure come from? How might the human mind discover the appropriate structure from experience? In this work, we enhance the influential model of structural discovery proposed by Kemp and Tenenbaum (2008) by leveraging generative flow networks. In doing so, we not only amortize inference times, but we also discovery a vast, novel distribution of structures given a datasets of entities and their features. Our moonshot goal is to develop a single GFlowNet than can sample structures for any subset of human knowledge.
Differentiable Linkage Kinematics in Geometric Algebra
Authors: Aditya Palaparthi, Alex Guerra, Ryan P. Adams
Institution: Princeton University
Date: In-Progress
Info: We are developing a differentiable linkage kinematics simulator using geometric algebra theory. In doing so, we not only greatly simplify the simulation logic of all types of linkage mechanisms but also simulate any type of 2D or 3D linkage mechanism.
Efficient Foundational Models for 3D PDEs
Authors: Udbhav Tripathi1, Navaneeth N.1, Aditya Palaparthi1,2, Jay Pathak1
Institution/Organization: 1Ansys, 2Princeton University
Date: In-Progress
Info: A foundational model for 3D PDEs has the potential to transform industries by accelerating 3D engineering simulations, reducing computational costs and enabling rapid prototyping and optimizations. This work aims to build upon the efficient 2D PDE foundational model Poseidon to address the challenge of foundational modeling for 3D PDEs. By leveraging pre-training and fine-tuning strategies, we extend the Poseidon paper’s capabilities to handle the added complexity of 3D PDEs. Our approach targets the unique demands of 3D problems, aiming to balance computational efficiency with generalizability.
Cleaning Dirty Meshes with Autoencoders
Authors: Aditya Palaparthi1,2, Jay Pathak1
Institution Organization: 1Ansys, 2Princeton University
Date: In-Progress
Info: Engineers involved in 3D scanning or engineering sim- ulation of CAD models often desire ways to correct er- rors in meshes, ranging from small holes to high polygon counts, in an efficient manner. Existing methods often in- volve a painstaking process for a mechanical design engi- neer to import an existing error-filled 3D mesh into existing 3D computer graphics software and manually correct with computationally inefficient algorithms. In this paper, I propose the first method to leverage artifical intelligence, specifically deep learning-based autoencoders, to efficiently error-correct 3D meshes at inference. I achieve high-quality results on correcting meshes with larger holes in the mesh, but often suffer on correcting meshes with small holes in the mesh.
Exploring Mechanical Linkage Design with Generative Flow Networks
Authors: Aditya Palaparthi, Alex Guerra, Ryan P. Adams
Institution: Princeton University
Date: May 2024 (Presentation @ Princeton)
Info:
Synthesizing optimal planar linkage mechanisms given only a traced path of motion is an intractable problem in mechanical engineering. Recently, machine learning solutions have been applied to tackle this linkage synthesis problem; however, no scalable solution has been found yet that can generate a set of optimal planar mechanical linkage designs, ranging in complexity, that best fit any type of generated path of motion by the user. In this work, we take advantage of the recently introduced flow-based generative model, generative flow networks (GFlowNets), and Laman graph theory, to explore the connection of single degree-of-freedom linkage mechanisms with GFlowNets. In doing so, we set the framework for a single generative model that can generate optimal mechanical linkage designs for a given path of motion.