A team has combined Bayesian optimization with synchrotron X-ray scattering to control the morphology of block copolymer blends in a closed loop. The approach, published in Nature, automatically adjusts blend composition to reach a target morphology without human intervention at each step.

Block copolymers self-assemble into nanostructures whose geometry depends on the composition and processing of the blend. Finding the recipe that yields a specific morphology —lamellae, cylinders or ordered spheres— usually relies on trial and error, a slow process when each characterization requires beam time at a synchrotron.

The closed-loop system described in the work chains three pieces: a Bayesian optimization algorithm that proposes the next composition based on previous results, the corresponding sample synthesis and processing, and characterization by X-ray scattering at an electron accelerator. The scattering signal is reduced to morphology metrics that feed the model back, which updates its prediction of the objective function and decides the next test.

Bayesian optimization is particularly suited to this problem because each measurement is costly and the parameter space is large: the model builds a probability distribution over possible outcomes and chooses the composition that maximizes information gained per experiment. The result is convergence toward the desired morphology in far fewer iterations than conventional systematic search.

The work illustrates how integrating machine learning with large-scale facility infrastructure can accelerate materials discovery. Beyond block copolymers, the same scheme could be applied to other self-assembled systems and to the processing optimization of functional polymers.