(154e) Designing Metamorphic Materials with a Machine-Learning Guided Genetic Algorithm

During natural selection, species evolve by generating large batches of variants and competing over numerous generations in search of optimized survival solutions. In the process, these systems learn the latent landscape and follow observed gradients toward optimization. The same idea can be largely adapted with machine learning algorithms and material discovery methods, designing automated systems that learn and optimize the design population.

Here, we demonstrate such a method for discovering designs for multi-stable metamorphic materials, DNA-responsive hydrogel architectures, that satisfy specific requirements for ready fabrication. Specifically, we ask whether it would be possible to fabricate metamorphic devices that, when presented with one of five or more different stimuli, can fold into five or more different functional shapes. We first develop a coarse-grained simulation that predicts the material conformations in response to different stimuli. We then train a machine learning model to score the predicted geometric outputs and provide fitness information regarding individual material designs. Finally, we build an evolutionary algorithm, which, taking feedback from the machine learning model, guides the evolution of the design population and optimizes autonomously. During the optimization process, we impose physical constraints on the mutation methods to limit design complexity and assure converged designs are fabricable under real-world conditions.

This work demonstrates that we are able to explore the vast design space efficiently and learn important trade-offs and relationships of the parameter landscape.