Using conditional normalizing flows to generate material placements in an optimized thermal composite

Abstract

We present an application of conditional normalizing flow neural networks (cFlows) to the problem of generating optimized topologies for burst-duty thermal mitigation, specifically targeting the placement of both linearly-conducting and phase-change materials (PCMs). Topological optimization naturally generates very large amounts of data. This enables us to acquire a high-resolution map of optimization space, accomplish cheap model generation, interpolate between points, and rapidly examine the statistics of the optimized topologies. The modeling is subject to two constraints: system mass (m) and maximum simulation end temperature (Tmax), i.e., our model takes these optimization constraints as inputs to predict the distribution of compatible topologies. The generated topologies are highly performant and realizable as verified by simulation; moreover, using the statistical properties of normalizing flows and an independent check against simulation, we show that the model is indeed sampling from the true distribution of possible qualified configurations.