Computational design automation and in-silico search of morphology

Lobo and Levin (2015) used an evolutionary algorithm to search for regulatory networks that account for experimental morphological phenotypes. A population of candidate networks evolved in parallel until one network produced zero error². The starting population held random networks with random parameters and random regulations among phenotypic products, a wound product, and the products of genetic and pharmacological perturbations². The search did not stop at the first fit. It continued until the complexity of the simplest zero-error network had not fallen for a set number of generations, because the mutation operators were biased toward simpler networks³.

Kriegman, Blackiston, Levin and Bongard (2020) described a pipeline for designing reconfigurable organisms built from generators and filters. They argued that this architecture lets elements be added, removed or reorganised. They suggested a further filter that would steer the search away from designs that cannot be realised physically⁴. Their build filter removes designs that are more than 50% muscle. The stated reasons were to leave design space for other specialised cells, such as sensory and memory cells, and to extend the lifetime of transferred designs, since contractile tissue has a much higher metabolic cost⁵.

Blackiston, Kriegman, Bongard and Levin (2023) reported that exploring design space in simulation filtered out billions of bad designs before any were built. This improved efficiency in the wetlab and reduced biowaste⁶. They also named a cost. Simulation is energy-intensive, and its carbon footprint grows with each second of simulation time. They called this arguably the most pressing ethical concern of xenobot research⁶. They proposed encoding design space so that gradients can be followed toward optimal solutions. If successful, moving from trial-and-error to gradient descent could cut processing time and carbon footprint by several orders of magnitude⁷.

Friston and Ao (2012) assumed that minimisation of free energy with respect to action and internal states conforms to a generalised gradient descent¹. Evolutionary search also appears inside models of development. Shreesha and Levin (2024) modelled embryos reaching a target morphology similar to the electric face bioelectric prepattern. They wrapped the developmental process in an evolutionary algorithm to compare effects on an evolutionary scale⁸. Pio-Lopez, Hartl and Levin (2025) trained a 16 x 16 cell NCA by evolutionary algorithms to perform morphogenesis of a smiley-face pattern, reminiscent of the craniofacial prepattern⁹.