Saroj UpretiSelf-driving labs for materials discovery
Control room

Closing the loop on block copolymer inverse design

A team-built closed loop in which a Gaussian-process optimizer drives a synchrotron beamline directly, finding the processing route to a named target morphology in 20–30 experiments instead of thousands.

Headline result
20–30experiments to hit the target, versus a grid scan of thousands
Period
2024 – presentActive
With
Lawrence Berkeley National Laboratory · Advanced Light Source
  • Bayesian optimization
  • gpCAM
  • GISAXS
  • Inverse design
  • ALS Beamline 7.3.3

The problem

Inverse design starts from the structure you need — here a 30 nm domain spacing, highly ordered, in a block copolymer blend — and asks which processing conditions produce it. Five processing knobs at ten settings each is 100,000 samples; no beamtime allocation covers a grid scan of that. The question: can the experiment choose what to run next, well enough to find the answer before the beamtime ends?

What we built

This was a collaborative effort with computational scientists and beamline staff at the Advanced Light Source. Together we connected three pieces end to end:

  1. Automated GISAXS reduction at beamline 7.3.3, turning raw frames into domain spacing and a measure of order before the next sample is in position.
  2. A Gaussian-process surrogate built on gpCAM, whose uncertainty tells the optimizer where it has measured versus where it is only predicting.
  3. An acquisition step that proposes the next condition and dispatches it to the robotic sample preparation.
sample librarydetector framesstructural descriptorsposterior + uncertaintynext conditionProcessRobotic thin-film fabricationMeasureSynchrotron GISAXSReduceDomain spacing, orderingModelGaussian process updateDecideAcquisition functionClosedloop20–30 iterations to target
One iteration of the closed loop. Each arrow was a separate integration problem: the reduction step has to keep pace with the beamline, and the model has to accept a measurement that may come back noisy or fail outright.

The result

For poly(styrene-b-2-vinylpyridine) (S2VP) blends, with a target of 30 ± 1 nm domain spacing and maximized order, the loop converged in 20 to 30 samples — roughly two orders of magnitude fewer than uniform sampling would need.

Reaching the target morphology

Best domain-spacing error found so far, against the number of samples measured. Lower is better.

Schematic
Closed-loop Bayesian optimizationUniform grid scan
Convergence toward the target domain spacingClosed-loop Bayesian optimization reaches the ±1 nm acceptance window after roughly 25 samples, while a uniform grid scan is still several nanometres away after 60.±1 nm acceptance window0102030405060036912Samples measuredDomain spacing error (nm)Closed loopGrid scan
View as table
Best domain-spacing error (nm) by number of samples measured. Schematic values.
SamplesClosed-loop Bayesian optimizationUniform grid scan
111.411.9
105.29.6
201.68.1
300.76.9
400.65.9
500.65.1
600.64.4

Nothing in the loop is specific to S2VP: it works over processing parameters and reduced structural descriptors, so it transfers to any system that can be made robotically and measured quickly. It is the foundation of the self-driving lab now being built.

Published in