Source-linked AI summary
Human-agent discovery of reconfigurable in-plane ferroelectric superdomain control
Yu Liu, Boris Slautin, Ching-Che Lin, Jaegyu Kim, Lane W. Martin, Sergei V. Kalinin
TL;DR
Exploratory instrument campaigns often lack predefined observables, actions, and objectives, limiting direct use of Bayesian optimization. This paper introduces SPARC, a human-agent framework for constructing and validating those elements, and applies it to reconfiguring ferroelectric superdomains, finding that alternating spatial polarity—not exact period matching—selects direction. The campaign also printed a UTK-shaped orientation pattern and exposed requirements for verified execution, validated observables, and robust control.
Problem
Exploratory experiments may require constructing sample variables, actions, and objectives during the campaign rather than defining them beforehand.
Method
SPARC pairs a coding agent and human operator with a microscope, shared notebook, persistent FINDINGS.md and PITFALLS.md files, preflight checks, and provenance.
Results
Alternating spatial polarity selected superdomain direction without requiring exact lattice-to-lamellar-period matching, and raster-plus-masked lattice writing produced a UTK pattern.
Takeaways & Limitations
Directional selection depended more on spatial polarity pattern and initial state than on exact commensurability or cumulative exposure.
Takeaways & Limitations
The main rewrite was demonstrated in one region, autonomous tests used few independent areas, retention lasted about an hour, and final states were not always pure orientations.
Abstract
from arXiv · showhide
Automated experimentation is most effective when the observables, available actions, and objective are defined before the experiment starts, as is the case for Bayesian optimization. However, in many exploratory experiments, the variables that describe the sample must be extracted from the data, new operations emerge during the experiments, and the instrument budget is too small to learn the problem by trials. Here we introduce the Scanning Probe Agentic Research Cycle (SPARC) framework, in which a coding agent and a human operator share one microscope, one notebook, and two persistent memory files. FINDINGS.md stores graded conclusions about the experiment, whereas PITFALLS.md records learned failure modes of analysis and instrument. We apply SPARC to reconfigure the in-plane superdomain direction of a (111)-oriented PbZr0.2Ti0.8O3 film. In an operator-supervised campaign, the agent reanalyzed earlier manual measurements and developed an oriented lattice of stationary bias pulses with alternating polarity to reconfigure the superdomain direction. In a subsequent agent-controlled campaign, PITFALLS.md entries were compiled into checks that validate a design before any write. The experiments showed that spatial polarity alternation, instead of the exact matching between the lattice and lamellar periods, determines directional selection. Combining a raster scan with a masked pulse lattice printed the letters UTK into the superdomain orientation. The campaign also identified practical requirements for agentic experimentation where physical verification of instrument execution, the conditions under which stored findings remain valid, validation of new observables on instrument data, and robust control protocols.
1. Introduction
Automated experimentation can adapt instrument operation, but Bayesian optimization assumes that the state, actions, and objective are already defined. SPARC addresses exploratory experiments where these elements must be constructed during the campaign.
- Automated instruments can combine programmable control, high-throughput analysis, and sequential decisions to adapt experiments as they run.
- Bayesian optimization has been adopted across synthesis, deposition, materials characterization, and scanning probe microscopy.
- The control problem concerns redistributing a thin film’s response among three in-plane polarization families rather than maximizing switched area.
- The state representation, action space, and comparison metric were unavailable initially and had to be developed during the experiment.
- The campaigns varied human-agent interaction, with the operator executing writes in Campaign 1 while the agent constructed states, hypotheses, controls, and programs.
2. Physics and data basis of the control problem
The paper frames superdomain control as an empirical mapping between crystallographically constrained states and probe actions, using lateral PFM to extract director populations. Because the same measured change can arise through multiple microscopic routes, the campaign focuses on reproducible redistribution among allowed director families.
- Physics and data basis: The physical representation constrains possible domain structures, while the experimental representation defines measurable states, probe effects, and prior knowledge for planning.
- Physics and data basis: A superdomain is a wide band of repeating nanodomain pairs whose in-plane director follows the lamellae, with three directors separated by 60° in lateral-PFM images.
- Physics and data basis: Changing director means redistribution among discrete superdomain families rather than continuous rotation of a single polarization vector.
- Physics and data basis: Lateral PFM measures only an in-plane projection, so observed director changes do not identify whether switching used ferroelastic steps, wall motion, or nucleation and growth.
- Control space: The empirical response function maps an applied tip action to a changed population vector without assuming a microscopic switching mechanism.
- Control space: Probe actions include stationary bias, biased raster or arbitrary trajectories, and spatial pulse arrays with control variables such as polarity, geometry, amplitude, waveform, and history.
- State extraction: Director populations were extracted from lateral-PFM stripe orientations using angular power spectra obtained by radial integration of image Fourier transforms.
- Experimental prior: Earlier fixed-protocol measurements supplied the campaign’s empirical starting point after rasters and continuous trajectories failed to reproducibly change director at the same location.
3. SPARC: the scanning probe agentic research cycle
SPARC combines an agent, operator, microscope, shared notebook, persistent scientific memory, executable checks, and provenance to support adaptive experiments whose representations and actions can change during a campaign.
- SPARC pairs a general-purpose coding agent with a microscope and operator, using a shared notebook, persistent memory, executable checks, and provenance.FINDINGS.md records graded conclusions and PITFALLS.md records analysis and hardware failure modes.
- SPARC workflow and review points: Each iteration reviews prior knowledge, forms a hypothesis, plans and validates an experiment, executes an action, interprets the readout, and reviews the result.The two campaigns differ in who owns the plan and result reviews.
- SPARC workflow and review points: SPARC allows the agent to replace state descriptors, redefine nulls, introduce action primitives, and retire control rules when data reject them.These revisions are traced to the data, code, and instrument state that motivated them.
- Agent-controlled operation: Executable checks converted selected pitfalls into preflight safeguards that could reject invalid or unsafe actions before execution.Checks included positive controls, scanner-range and bias limits, treatment footprints, and per-iteration action budgets.
- Building the experimental state: The agent constructed and refined a quantitative domain descriptor by comparing orientation estimators, fitting three direction families, and registering successive images.Estimator validation used synthetic stripe patterns with known ground truth before applying the descriptor to experimental images.
4. Campaign 1: Operator-controlled discovery of a reconfigurable primitive
In Campaign 1, the agent reanalyzed failed protocols and developed stationary, alternating-polarity pulse lattices that could select and partially redirect in-plane superdomain populations under operator control.
- Campaign 1 design: Campaign 1 used operator-executed writes while the agent analyzed evolving states, proposed instrument actions, and interpreted readouts.The immediate goals were selecting one allowed direction and redirecting a written region toward another.
- Action discovery: Code reanalysis revealed that nominally different polarities could generate identical files and that dense trajectories reversed bias within a lamellar period.These findings motivated spatially resolved stationary pulse lattices with alternating polarity.
- Reconfiguration experiment: In R6, write A increased P1 power near 4° by +0.392, while the independently written P3 increased by +0.577 and untreated P2 changed by −0.006.The untreated movement was within the run-specific floor of 0.018.
- Reconfiguration experiment: During write B, P1 power near 4° decreased by -0.287 while power near 124° increased by +0.149.The retention control relaxed by only -0.076, and write B increased untreated P2 power near 124° by +0.111.
- Interpretation and validation: The final P1 state remained a mixture of two director families rather than a complete reorientation.Because dominant-angle labels can be unstable when families have similar weights, the claim relied on controlled population changes.
- Interpretation and validation: The agent validated orientation readouts by comparing Fourier, edge-detector, and structure-tensor methods on synthetic stripe patterns with known ground truth.The structure tensor performed best on the synthetic data, while the edge detector failed for finer stripes.
5. Campaign 2: agent-controlled testing and composition of control rules
Campaign 2 established that alternating-polarity pulse lattices selected superdomain orientations without requiring exact period matching, enabled reversible but non-deterministic reconfiguration, and combined with raster writing to print UTK.
- Autonomous hypothesis tests: Alternating-polarity lattices increased the targeted orientation across repeat distances of one, two, four, and eight lamellar periods, so exact period matching was unnecessary.Uniform-polarity arrays produced no directional change above baseline variation, while alternating-polarity controls did.
- Autonomous hypothesis tests: Directional selection remained observable at 70% of the exposure threshold estimated from Campaign 1, and final orientation populations stayed similar across an approximately 2.2-fold exposure range.Both findings superseded the proposed exposure threshold as a necessary condition.
- Set, reverse, and hold the orientation state: Seven of eight valid trials ended with the targeted orientation dominant, demonstrating reversible selection without deterministic control.One of nine trials was excluded because its reference control failed.
- Set, reverse, and hold the orientation state: The written orientation persisted for 34 min during repeated imaging, although longer measurements are needed to establish long-term retention.Successive writes could change the dominant orientation and return it toward 64°, while an untreated reference remained unchanged.
- Composing the rules: printing UTK into the superdomain orientation: A 62° raster prepared the background, while a masked alternating-polarity lattice commanded to 2° printed the letters UTK into the superdomain orientation.The lattice contained equal numbers of positive and negative pulse sites and was applied only along the letter strokes.
- Composing the rules: printing UTK into the superdomain orientation: Relative to spaces between the strokes, the target population increased by 0.30 in U, 0.42 in T, and 0.52 in K.Each letter was approximately 3 μm across because the analysis window had to span several lamellar periods.
6. Discussion
The campaigns established a hierarchy of controllable variables and a workflow in which human expertise, agent automation, persistent memory, and executable checks jointly supported closed-loop experimentation. The results also expose limits in both the physical evidence and agent-authored control software.
- Materials rules for in-plane superdomain control: Polarity alternation was necessary under tested conditions, whereas exact matching between pulse-array and lamellar periods was not.The resulting protocol used 1 s, ±10 V pulses in a rotated square lattice with alternating row polarity.
- What the framework did: The agent converted operator feedback into persistent findings, recorded analysis and hardware failure modes, and compiled selected pitfalls into executable design checks.Across nine closed-loop cycles, it halted eight without operator intervention and revised rules based on new measurements.
- What the framework did: Control software remained a weakness because six of eight launches failed from faults that static checks could have detected.This finding complements the framework’s successful use of executable checks while showing that the checks did not cover all tooling failures.
- Defining and grounding the experiment: The experiment was not yet suitable for unrestricted optimization because the state, actions, objectives, preconditions, provenance, controls, read-backs, uncertainty, and validity ranges required explicit grounding.The paper positions Bayesian optimization as appropriate only after the exploratory campaign reduces the problem to defined parameters and a reproducible response range.
- Human-agent complementarity: Human and agent contributions were complementary: the operator supplied instrument-specific knowledge and physical insight, while the agent implemented analyses, controls, records, and later microscope commands.The shared infrastructure changed between campaigns through persistent memory, revised analysis code, explicit state and action definitions, and controlled microscope access.
- Limitations and next experiments: The physical evidence remains limited by sparse regions and independent areas, short retention measurements, mixed final states, coupled write geometries, one cantilever orientation, and an altered out-of-plane response.The authors call for additional replication, matched geometries, longer retention, vector PFM or sample rotation, and measurements separating switching, wall motion, and nucleation.
7. Summary
SPARC supports exploratory microscopy before observables, operations, and objectives are fixed by coupling a coding agent with human supervision and persistent experiment records. Applied to a ferroelectric film, it developed and tested pulse-lattice and raster protocols for reconfiguring in-plane superdomain orientations while revealing safeguards needed for direct instrument control.
- SPARC couples human operators with a coding agent that analyzes data, proposes experiments, and later controls the microscope after safeguards are implemented.FINDINGS.md stores conclusions with confidence and applicable conditions, while PITFALLS.md stores failures and checks read during subsequent planning.
- Campaign 1 used earlier PFM measurements to develop a quantitative superdomain descriptor and an alternating-polarity pulse lattice, with the operator retaining physical control.
- The pulse lattice preferentially increased a targeted crystallographic orientation and redirected previously written regions toward another allowed orientation.Within the investigated range, polarity alternation was important for directional selection, whereas exact period matching and exposure threshold were not required.
- A bipolar raster aligned previously unwritten regions along the scan direction, and combining raster preparation with masked pulse-lattice writing produced a UTK-shaped pattern.
- Direct instrument control required separate checks for command execution, sample response, control measurements, data provenance, and image-derived descriptor validity.Some safeguards remained specific to the microscope and measurement protocol, although SPARC provides a framework for adapting safeguards to other instruments with prior knowledge and human supervision.
- SPARC shortens the experimental loop through data analysis, record-keeping, and measurement selection while leaving scientific objectives, physical interpretation, and problem reformulation to the researcher.Once observables, control space, and objectives stabilize, the problem can be transferred to conventional sequential optimization.
8. Materials and methods
The methods combine a PZTO/LSMO heterostructure, stationary biased sites arranged on a rotated square lattice, and a software-controlled campaign architecture. Instrument operation used persistent state, constrained candidate writes, exclusive access, and stop controls, while implementation details included a coding agent and shared records.
- The model system was a 150 nm PbZr0.2Ti0.8O3 layer on 30 nm La0.67Sr0.33MnO3 over (111)-oriented SrTiO3, grown by pulsed-laser deposition.A KrF excimer laser with λ = 248 nm ablated ceramic targets with nominal compositions Pb1.2Zr0.2Ti0.8O3 and La0.67Sr0.33MnO3.
- The pulse lattice used stationary biased sites on a square grid rotated to the commanded director, with representative 280 nm lamellae giving 140 nm row spacing.Campaign 1 used 14 x 14 or 12 x 12 sites per addressed 2 μm panel, with nominal 10 V, 1 s pulses and alternating row polarity.
- The same coding-agent model and tool set ran both campaigns, with persistent state stored in a shared notebook, FINDINGS.md, PITFALLS.md, and a Campaign 2 JSON state file.Campaign 1 was human mediated, with the operator remaining present during agent-generated analyses, experiment cells, and write programs.
- The implementation also drew on prior work in automated microscopy, reward-based image analysis, and coding agents for scientific discovery.
- Each direct-actuation iteration followed propose, preflight, place, write, read, analyze, log, and decide steps through the AESPM control interface.An exclusive lock prevented simultaneous instrument access, a file-level stop command could halt the loop before new actions, and candidate writes were constrained by scanner, bias, and exposure limits.