Workshop: AI4S: 8th Workshop on Artificial Intelligence and Machine Learning for Scientific Applications
Authors: Leslie Horace (Georgia Institute of Technology) and Vanessa Job, Sharmistha Chakrabarti, and Nathan DeBardeleben (Los Alamos National Laboratory)
Abstract: Generative inverse design for scientific simulations addresses a many-to-one problem: generating distinct inputs whose forward simulations approximate a target output. Our prior work generated image-based initial conditions directly, but simulations often construct initial conditions at runtime from configuration variables. We shift the task upstream by generating variables whose effects propagate through intermediate representations before producing the final output. We use a deterministic heat-diffusion proxy with Cosine Phase and Perlin Noise geometry functions that create diverse material-layout patterns. Conditional normalizing flows generate input variables conditioned only on observed final states, and the resulting simulated outputs are evaluated against the corresponding targets. Variable-inclusion ablations isolate geometry and thermal variables, while posterior-factorization ablations compare full-joint, chained-joint, and disjoint decompositions. Results show error arises primarily when modeling geometry variables, whereas posterior factorization provides little improvement over baselines. These findings suggest the final outputs may not retain enough information about upstream variables and intermediate representations.
Back to AI4S: 8th Workshop on Artificial Intelligence and Machine Learning for Scientific Applications Archive Listing
Back to Full Workshop Archive Listing