Publications

Code, notebooks, and compact datasets behind my publications on autonomous and machine-learning-driven scanning probe microscopy.

Every folder here is a self-contained release for one paper: the notebooks run top to bottom (most of them on a free Colab CPU runtime), the data needed to reproduce the published numbers is either included or pointed at, and the instrument-facing code is published as it actually ran.

Richard (Yu) Liu · University of Tennessee, Knoxville · Google Scholar · yliu206@utk.edu / yu93liu@gmail.com

The microscope control layer used by most of this work is a separate package: aespm.

---

Table of contents

Project	What it does	Paper	Status
SPARC-PZTO111	A coding agent and an operator share one microscope and two persistent memory files for a full campaign — ending with UTK written into ferroelectric domain orientation	Liu et al., Digital Discovery (2026)	submitted
DT-SPM	Descriptor-aligned digital twins of an SPM: physics simulator + learned corrections that predict scan quality on unseen conditions	Liu et al., Digital Discovery (2026)	submitted
GA-DKL	Evolutionary search in Fourier space + uncertainty-aware DKL discovers tip-bias waveforms that electrically de-age a ferroelectric film	arXiv:2606.13859	preprint
MOBO-DKL	Multi-objective deep kernel learning over 427 destructive PFM measurements; pulse-dose Pareto front for polarization switching	arXiv:2506.08073	under review
Combinatorial-library SPM	Noise-aware and prior-informed Bayesian optimization for automated PFM across composition-spread libraries	J. Appl. Phys. 140, 084101 (2026)	published
Pareto-optimal experimentation	Human-guided multi-objective Bayesian optimization: how reference points and reward weights steer an autonomous experiment	Nano Lett. 26, 441–447 (2026)	published
Tutorials and simulators	Hardware-free notebooks for MOBO, DKL, and the automated-platform workflow	—	—

---

SPARC-PZTO111 — human–agent discovery of in-plane superdomain control

SPARC-PZTO111/ — Scanning Probe Agentic Research Cycle.

SPARC pairs a coding agent with one microscope and one operator for the length of a campaign. The experiment asks which tip operations select the in-plane superdomain direction of a (111)-oriented PbZr_0.2Ti_0.8O₃ film, where the three allowed directions are related by symmetry, so a uniform field cannot choose between them.

Highlights

• The memory files are the artefact. memory/FINDINGS.md carries a reliability grade on every claim about the sample; memory/PITFALLS.md records how the analysis has previously gone wrong. Both are published verbatim, as the campaign left them.
• Pitfalls became pre-flight checks. Selected PITFALLS entries were compiled into checks that refuse a design before any voltage reaches the tip — scanner range, bias ceiling, declared footprint, write budget (PITFALLS.md §8).
• Two campaigns, published as they ran. Campaign 1 was operator-supervised (R1–R7); campaign 2 was agent-driven (IT1–IT10b), one script per iteration in campaign2/iterations/ with the recorded consoles beside them. The run ends with the letters UTK written into domain orientation.
• Every file sent to the instrument is in the repo (write_files/), along with SHA-256 manifests for all 676 raw frames and the coded interaction record behind Supplementary Section S12.2.
• setup_workspace.py rebuilds the flat directory the analysis scripts were written for, by hard link, so the published layout stays readable without duplicating data on disk.

Cite

Y. Liu, B. Slautin, C.-C. Lin, J. Kim, L. W. Martin and S. V. Kalinin, Human-agent discovery of reconfigurable in-plane ferroelectric superdomain control, submitted to Digital Discovery (2026).

See CITATION.cff. Code MIT; data, memory files and records CC BY 4.0.

---

DT-SPM — coupled digital twins for predictive and autonomous microscopy

DT-SPM/ — descriptor-aligned digital twins for scanning probe microscopy.

Measured force–distance curves and a grid-scan archive are reduced to a locked vocabulary of operational descriptors — 18 FD descriptors plus five scan-quality descriptors (Q_align, Q_stab, Q_grad, Q_safety, Q_range). A physics-informed neural encoder recovers the FD descriptors from raw amplitude/phase curves; the descriptors re-parameterise a deterministic feedback-scanner simulator; compact CNN-LSTM corrections close the remaining gap to experiment.

Highlights

• The learned correction is what makes the twin predictive. On held-out Tap-300/AlScN conditions the Q_align median error drops 10.4 nm → 0.6 nm and the rank correlation with experiment rises ρ = +0.23 → +0.67. Line RMSE on the calibration grating drops 48 nm → 20 nm.
• Two probe/sample systems: Tap-300 / AlScN (900 scan conditions) and Multi-75 / calibration grating.
• Reproduce the paper in minutes. DT-SPM_reproduce_paper.ipynb verifies every manuscript number and regenerates all 14 figures in ~1–3 min on a CPU.
• Run the twin end-to-end. DT-SPM_full_digital_twin.ipynb runs the complete pipeline (FD physics → descriptors → PINN → scanner → CNN-LSTM) and ships a transfer playbook for adapting it to new measurements. No GPU needed.
• NOTEBOOKS.md documents every section, knob, and expected output.

Cite

Y. Liu, B. Slautin, I. Mercer, J.-P. Maria and S. V. Kalinin, From Closed-Loop Optimization to Open Decision Making: Coupled Digital Twins for Predictive and Autonomous Microscopy, submitted to Digital Discovery (2026).

Related: S. V. Kalinin, Y. Liu, B. Slautin et al., Dual Digital Twins for Experimental Planning in Automated Laboratories (2026).

---

GA-DKL — closed-loop discovery of out-of-distribution processing protocols

GA-DKL/

Many functional properties are governed not only by composition and structure but by history — the time-dependent protocol that brings a material to its operating state. Here the protocol is a scanning-probe tip-bias waveform and the reward is the measured change in effective nonlinear electromechanical response (ENL) before and after it.

Highlights

• VAE initialization manifold. A 1D convolutional VAE trained on 14 experimentally common seed waveforms; decoding a 21 × 21 latent grid gives a continuous family of realistic seeds.
• GA in Fourier space. Each waveform is 33 Fourier coefficients (DC + 16 harmonics); tournament selection, BLX-α crossover, and Gaussian mutation generate ~1,000 out-of-distribution candidates per generation, with amplitude bounds enforced at synthesis.
• DKL surrogate ranks before you measure. A 1D CNN feature extractor with an exact GP head scores candidates by upper confidence bound, so only a small informative batch reaches the microscope.
• The result is physical. On PZT thin films the campaign converges on temporally structured, multi-harmonic waveforms that enhance nonlinearity by selectively depinning weakly pinned domain walls — in effect electrically de-aging the film.
• Runs without hardware. The workflow notebook carries a self-contained appendix with a synthetic objective; swap evaluate_waveform() for your own instrument call to go live.

Cite

Y. Liu, S. Udovenko, C.-C. Lin, J. Kim, L. W. Martin, S. Trolier-McKinstry and S. V. Kalinin, Closed-loop discovery of out-of-distribution processing protocols by evolutionary search and uncertainty-aware learning, arXiv:2606.13859 (2026). arXiv

---

MOBO-DKL — reward-guided multi-objective deep kernel learning of polarization switching

MOBO-DKL/ —

Autonomous physics discovery of polarization switching mechanisms in ferroelectrics. The release ships 427 destructive structure–response measurements on two pre-poled PZTO-(111) areas, together with the analysis that turns them into a decision rule.

Highlights

• Switched area is audited, not asserted. The notebook independently recomputes the 45° phase-change and amplitude-gated masks — including the global 180° flip guard and the center-localized connected-component rule — and all 427 areas must reproduce the released values exactly.
• A pulse-dose Pareto front. The empirical front minimizes the dose proxy |V|·t while maximizing switched in-plane area; switching probability and the voltage–dwell efficiency map fall out of the same table.
• Structure predicts response. Fifteen structural descriptors extracted before each pulse, plus a six-dimensional convolutional VAE that never sees voltage, dwell, switched area or efficiency.
• Replay-mode active learning. Independent reward-specific CNN-GP surrogates and qLogEHVI recommend a held-out recorded candidate, testing the acquisition logic without a microscope. A typed adapter defines the four operations a lab must implement to go live.
• Data is versioned and checksummed — data/DATA_DICTIONARY.md and data/SHA256SUMS, arrays load with allow_pickle=False.

Cite

Y. Liu, U. Pratiush, K. Barakati, H. Funakubo, C.-C. Lin, J. Kim, L. W. Martin and S. V. Kalinin, Domain Switching on the Pareto Front: Multi-Objective Deep Kernel Learning in Automated Piezoresponse Force Microscopy, arXiv:2506.08073 (2025). arXiv

Published as Autonomous Physics Discovery of Polarization Switching Mechanisms in Ferroelectrics via Reward-Guided Multi-Objective Deep Learning (under review, 2026). Development notebooks for this study live at the repository root: MOBO_DKL_RL_dev_v2(qEHVI).ipynb and Single_DKL_RL_dev1.ipynb.

---

Combinatorial-library SPM — Gaussian-process-guided exploration

BO_simulation_for_automated_exploration_of_combi_library.ipynb —

A fully automated SPM workflow for ferroelectric combinatorial libraries — stage motion, probe engagement, in-contact tuning, imaging, DART-PFM spectroscopy, and the choice of the next measurement site all proceed without human input. Demonstrated on Sm_xBi_1−xFeO₃ and Zn_xMg_1−xO libraries.

Highlights

• The notebook replays a 20-point grid search on the Sm-BFO library (mean and measured variance of the loop height at each site), interpolates it, and uses it as the ground truth for active-learning simulations you can rerun in a browser.
• Noise-aware BO (nBO): a second GP models the measurement noise, so the acquisition function can avoid sites that are only interesting because they are noisy.
• Prior-informed BO (vBO): user-defined priors encode what the experimentalist already believes about the composition–property trend, and the notebook shows what that buys and what it costs.
• Seeding, exploration budget, and acquisition are all exposed as knobs — 5 seeds + 20 steps in the published run.

Cite

Y. Liu, R. Pant, I. Takeuchi, R. J. Spurling, J.-P. Maria, M. Ziatdinov and S. V. Kalinin, Automated scanning probe microscopy of combinatorial ferroelectric libraries: Gaussian-process-guided exploration and noise-aware experiment planning, J. Appl. Phys. 140, 084101 (2026). doi:10.1063/5.0341018

The multi-objective, multi-modal extension of the same platform — MOBO_simulation_v2.ipynb — builds the MOBO loop as four replaceable pieces (rewards_mobo(), measure(), generate_seed_mobo(), step()) so the same workflow drives conductance, film quality, ferroelectric response, or purely geometric descriptors.

Y. Liu, A. Raghavan, U. Pratiush, M. Ziatdinov, C.-Y. Lee, R. Pant, I. Takeuchi et al., Automated Materials Discovery Platform Realized: Scanning Probe Microscopy of Combinatorial Libraries, arXiv:2412.18067 (2024). arXiv

---

Pareto-optimal experimentation — human-guided MOBO on an SPM simulator

AC MOBO based on SPM simulator_v5.ipynb · Colab-ready copy: AC_MOBO_based_on_SPM_simulator_v5.ipynb

What does a human actually control in a multi-objective autonomous experiment? Not the individual measurements — the shape of the trade-off. This study runs the whole argument on the spmsimu simulator, so every claim is reproducible without a microscope.

Highlights

• Builds single- and multi-reward functions on simulated AC-mode SPM data, then does a full grid search of the reward landscape as ground truth to check the optimizer against.
• Shows how the reference point and the reward weights — the two things a human really sets — move the discovered Pareto front.
• Head-to-head comparison with single-task GP / scalarized BO on the same problem and the same budget.
• ~120 cells, hardware-free, runs on a free Colab CPU runtime.

Cite

Y. Liu and S. V. Kalinin, Pareto-Optimal Experimentation: Human-Guided Multi-Objective Bayesian Optimization in Scanning Probe Microscopy, Nano Letters 26(1), 441–447 (2026). doi:10.1021/acs.nanolett.5c05373 · preprint: The Power of the Pareto Front, arXiv:2504.06525

---

Tutorials and simulators

Hardware-free notebooks at the repository root. Start here if you want the method rather than a specific paper.

Notebook	What it teaches	Colab
MOBO Tutorial_v2.ipynb	Multi-objective BO from scratch with BoTorch/GPyTorch: defining the problem and parameter space, shifting the reference point, weighting rewards
MOBO_simulation_v2.ipynb	The reusable MOBO experiment scaffold (rewards_mobo / measure / generate_seed_mobo / step) used by the automated-platform paper
MOBO_DKL_RL_dev_v2(qEHVI).ipynb	Deep kernel learning inside a multi-objective loop with qEHVI, on simulated then grid-measured data
Single_DKL_RL_dev1.ipynb	The single-objective DKL baseline behind the domain-switching study

Built with BoTorch and GPyTorch; instrument control via aespm; AC-mode simulation via spmsimu.

---

Getting started

Each project folder carries its own requirements.txt.

git clone https://github.com/RichardLiuCoding/Publications.git
cd Publications/<project>
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
jupyter lab

Python 3.10+ is recommended. A GPU accelerates VAE and DKL training but is not required by any notebook here — the reproduce-the-paper paths are all CPU-only.

If you fork this repo: copy it out of any cloud-synced folder before running git. Dropbox, Google Drive, and OneDrive rewrite files under .git/ while git is using them.

Citing

Please cite the paper for the folder you used — the citation is in each section above and in the folder's own README.md. If you use the optimization stack, also cite BoTorch and GPyTorch as listed in the corresponding Methods section.

Licence

Licensing is per project. SPARC-PZTO111 is MIT for code and CC BY 4.0 for data, memory files, derived results, and records (LICENSE, LICENSE-DATA). Other folders follow the terms stated in their own README.md.

Contact

Open an issue, or email yliu206@utk.edu · yu93liu@gmail.com. Full publication list: Google Scholar.