Session: 13-22-02: CONCAM Distinguished Lectures on Computational Mechanics II

Paper Number: 174009

Constitutive Modeling in the Era of Ai. Part Ii: Guided Generations of Microstructures With Designated Properties 

The second part of this talk is about inverse problems enabled by generative AIs. Given a set of desired properties, we would like to inversely design microstructures that fit the descriptions we provided. In this study, we consider two-phase composites where crystals are distributed in the host polymer matrix, e.g., a polymer-bonded explosive. The dynamic behaviors of polymer-bonded explosives (PBX) are dominated by the morphology of individual particles and the topology of microstructures. Over the past decades,  a growing body of work has established many critical multiscale relationships between key microstructural attributes. However, due to the cost and labor required to conduct shock experiments in the PBX with energetic crystals, it is not always feasible to use experiments alone to acquire the necessary data to construct the response surface between performance and controlling parameters. Presumably, it is possible to directly train latent diffusion models to generate realistic-looking microstructures. However, unlike most computer vision tasks that aid only in realistic-looking images, the microstructures we generated are intended for high-fidelity tasks, such as extending material databases, quantifying uncertainty and sensitivities, and other simulation tasks that require realistic mechanical properties. In the case of polymer-bonded explosives (PBXs) where hot spots found in the crystals often interact, preserving the topology features of the microstructures could be critical to generate microstructures with realistic material properties. In other words, a more direct control over how particles should be distributed in the polymer matrix is needed. Given the need to preserve topological features while generating microstructures with crystals of realistic geometries, we introduce a staggered generative AI that breaks down the generation tasks of PBX into two sub-tasks handled by two coupled generative algorithms, i.e., (1) a GraphRNN autoregresive model that generates the topology of synthetic polymer-bonded composites, where features and locations of the crystal inclusions are prescribed and (2) a latent diffusion model that generates individual crystal inclusions with conditions introduced to fit the features prescribed by the GraphRNN. This approach enables us to consider each crystal as a simply-connected body embedded in three-dimensional space, hence avoiding the otherwise necessary but challenging task of generating manifolds of complex topology. To generate PBX microstructures with realistic crystal topology, we sample sub-domains of segmented micro-CT images of a mock PBX with Idoxuridine (IDOX) to constitute a data set. Each sub-domain is represented by a node-weighted graph with the position, size, and shape descriptors stored as node features. The topological data set is then used to train a graph recurrent neural network algorithm (GraphRNN) that learns to sequentially add nodes and edges to form a weighted graph that describes the IDOX network, consisting of nodes with features connected by edges. With the global topology prescribed, the individual crystals are generated by a conditional denoising diffusion probabilistic model that takes the node features as inputs and creates individual crystals that fit the prescribed node features generated from the GraphRNN. Numerical examples are provided to analyze the similarity between the actual and synthetic microstructures. The synthetic PBX microstructures are provided for third-party validation.

Presenting Author: Waiching Sun Columbia University

Presenting Author Biography: Dr. Steve Sun is an associate professor of civil engineering and engineering mechanics at Columbia University. He received his PhD from Northwestern in 2011. From 2011 to 2013, He worked as a research engineer at Sandia National Laboratories. Sun’s research focuses on computational mechanics and scientific machine learning for material modeling. He received several awards, including the Walter Huber Prize and da Vinci Award from ASCE, the John Argyris Award from IACM, and the CAREER award from NSF, the Army, and the Air Force. Since April 1st, 2025, he has become an editor of the International Journal for Numerical Methods in Engineering.

Authors:

Waiching Sun Columbia University