AUTOPROBE: Autonomous Probing in Microscopy & Code

• AUTOPROBE is an autonomous probing paradigm that enables systems to inspect their operating state, decide on the adequacy of conditions, and perform corrective actions in microscopy and code evaluation.
• In scanning probe microscopy, it uses a CNN-based classification to diagnose tip health, achieving image-level precision of 97% and over 99% with majority voting for reliable measurements.
• For LLM code evaluation, AUTOPROBE dynamically selects informative internal representations, improving correctness assessment by up to 111% in functionality and outperforming fixed-layer approaches.

AUTOPROBE denotes an autonomous probing paradigm in which an experimental or computational system inspects its own operating state, decides whether the current condition is acceptable for the task, and then either proceeds, redirects the measurement, or performs corrective action. In scanning probe microscopy, the term is associated with closed-loop automation for probe-health monitoring, drift compensation, defect avoidance, atomic-site localization, and in-situ recovery of usable probe states (Rashidi et al., 2018, Diao et al., 2024). In a separate usage, AUTOPROBE is also the name of a model-agnostic white-box method for assessing the correctness of LLM-generated code through dynamic selection of informative internal representations (Vu et al., 3 Oct 2025).

1. Scope and principal meanings

Across the cited literature, AUTOPROBE appears in two distinct but technically related senses: as a microscopy autonomy concept centered on self-correcting probe operation, and as the proper name of a probing method over neural internal states (Rashidi et al., 2018, Vu et al., 3 Oct 2025).

Usage	Domain	Defining capability
Closed-loop autonomous probe conditioning	STM/SPM	Detect degraded probe state and trigger in-situ tip restoration
AI-equipped self-driving measurement	Room-temperature SPM	Judge tip/sample condition, compensate drift, choose sites, and keep measuring
Novelty-driven discovery workflow	Automated microscopy	Search for diverse target-space responses rather than only optimize known objectives
Dynamic internal-state probing	Code LLM evaluation	Select informative hidden states for correctness assessment

In the microscopy literature, the central motivation is operational fragility. Atomic-scale STM/SPM depends on an atomically sharp tip apex, ideally with a single dominant atom at the tip end. When the apex becomes blunt or multi-apexed, the microscope loses atomic resolution, produces secondary imaging features, and, in atomic fabrication, incurs large positioning errors that undermine atom-by-atom writing (Rashidi et al., 2018). Room-temperature operation adds thermal drift, probe-apex variation, and the recurring need to find atomically clean, defect-free regions before spectroscopy or manipulation can proceed (Diao et al., 2024).

A recurrent misconception is that such autonomy necessarily implies reinforcement learning or online adaptation. The 2018 STM implementation is explicitly not of that kind: the autonomy comes from integrating a pre-trained classifier with microscope control logic and a deterministic actuation rule, rather than from reinforcement or online learning (Rashidi et al., 2018).

2. STM tip-quality diagnosis and in-situ conditioning

A canonical AUTOPROBE implementation was presented in “Autonomous Scanning Probe Microscopy in-situ Tip Conditioning through Machine Learning” (Rashidi et al., 2018). The system is a closed-loop “sense–decide–act” routine for STM on hydrogen-terminated Si(100), using isolated surface dangling bonds (DBs) as diagnostic test objects. At $-1.8$ V and 50 pA, isolated DBs appear as bright protrusions. A sharp tip renders them as clean, isolated features with the expected symmetry and appearance, whereas a degraded or “double tip” alters the apparent DB shape, introduces secondary contrast, or otherwise distorts the feature because the tunnel current is effectively sampled by more than one apex.

The training corpus comprised roughly 3500 STM images of isolated DBs assembled from five years of data taken on two microscopes. Each DB image was cropped to a $5.6 \times 5.6$ nm$^2$ region and resized to $28 \times 28$ pixels. Every image was manually labeled as sharp-tip or double-tip, and the dataset was enlarged by rotations through $90^\circ$ increments and mirroring, yielding an 8× augmentation. Several classifiers were compared on the same dataset: k-nearest neighbors, random forest classifier, support vector machine, a fully connected neural network, and a convolutional neural network. The reported precision values were approximately 0.84 for KNN, 0.89 for random forest, 0.88 for SVM, 0.78 for the fully connected neural network, and 0.97 for the CNN (Rashidi et al., 2018).

The CNN selected for automation used two convolution layers with 30 and 40 kernels, each $5\times5$ with stride 1, followed by ReLU activations, a $2\times2$ max-pooling layer with stride 2, flattening, a fully connected 128-node ReLU layer, and a final 2-node softmax output layer. The output classes were encoded as “0” for sharp tip and “1” for double tip. Training used Adam with learning rate $10^{-4}$ and categorical cross-entropy loss. At the single-image level, the CNN correctly identified degraded tips in 97% of test cases.

Operational robustness was increased by classifying multiple DBs extracted from a larger frame and then applying majority voting,

$\hat{y} = \mathrm{mode}\{y_1, y_2, \dots, y_n\},$

where $y_i \in \{0,1\}$ are the per-DB predictions. This improved the effective accuracy to beyond 99%. The voting step suppresses false positives caused by a nearby defect or an anomalous DB image, which is important because the paper notes that a nearby defect can itself be misread as double-tip-like contrast (Rashidi et al., 2018).

The control loop is intentionally simple. A user selects an assessment frame and a preselected spot for safe conditioning. The STM acquires the frame, the software locates dangling bonds, the CNN classifies each candidate DB, and majority vote determines whether the tip is acceptable. If the vote indicates a double tip, the system initiates tip conditioning by controlled indentation into the sample, then reimages and reclassifies. The loop stops only when the majority vote returns “sharp tip.” In one example, a visibly degraded tip required four conditioning steps; in another, during atom-by-atom wire fabrication, the system detected unexpected degradation, performed three tip-conditioning events, confirmed recovery, and resumed fabrication (Rashidi et al., 2018).

3. Room-temperature AI-SPM as a broader AUTOPROBE implementation

“AI-equipped scanning probe microscopy for autonomous site-specific atomic-level characterization at room temperature” extends the AUTOPROBE idea from tip-health restoration to a broader self-driving measurement platform (Diao et al., 2024). The system combines conventional SPM hardware with deep-learning inference, thermal-drift correction, and automatic probe conditioning, so that the microscope can inspect the current scan, decide what is usable, relocate to the next target, and keep measuring when tip state or surface conditions change.

The architecture is split between an AI inference side and a scan-control side. The inference module receives topographic images in real time and classifies tip and sample condition, detects adsorbates and defects, and localizes atomic sites. On Si(111)-(7×7), three CNNs are used. Net1 classifies overall tip/sample condition into 11 labels $5.6 \times 5.6$0 and decides whether the state is suitable for site-specific measurement; it also triggers tip shaping when the probe is not good enough. Net2 detects adsorbates, distinguishing moving adsorbates $5.6 \times 5.6$1 from stationary adsorbates $5.6 \times 5.6$2. Net3 identifies surface geometry and atomic positions, classifying half-unit cells as faulted or unfaulted ($5.6 \times 5.6$3, $5.6 \times 5.6$4) and localizing keypoints $5.6 \times 5.6$5 corresponding to adatoms and corner holes (Diao et al., 2024).

The control logic is hierarchical. Net1 first decides whether the current tip/surface condition is “good” enough; Net2 and Net3 then determine whether the region contains defects or atomically clean target sites; the scan module either acquires spectroscopy/manipulation data or avoids contaminated regions and steps. Thermal drift is handled by a feedforward correction scheme based on feature-point matching between consecutive images. Drift velocity is estimated from image-to-image shifts and scan timing, used to offset the scan position in real time, iterated until the measured drift drops below a 0.2 nm threshold, and updated every 10 minutes after compensation. Tip-apex variation is monitored through Net1. If the CNN judges the tip state non-optimal, the system performs automated tip shaping by bringing the tip closer to the surface and changing the bias/current conditions; a representative poke is 0.9 nm toward the surface at 1.5 V and 200 pA, and if no change occurs the next poke moves 0.15 nm closer (Diao et al., 2024).

The quantitative performance is unusually explicit. Net1 has average recall 0.93, and binary good-versus-bad tip judgment reaches 0.98 recall. Net2 has average recall 0.92 at an IoU threshold of 0.7. Net3 achieves mAP above 0.98, with mAP at $5.6 \times 5.6$6 equal to 0.91, and unit-cell classification recall approaches 1.0. In autonomous imaging, the system acquired 45 consecutive images over different regions while compensating drift and bypassing impurities, defects, and steps. In the main room-temperature STS demonstration, it acquired 324 I–V curves on four adatom types; over 58 hours, it produced 695 topographic STM images and 2832 I–V curves in total (Diao et al., 2024).

The same paper also sets an important boundary on claims of robustness. Tip changes during I–V sweeps remained a real issue even with automation: preliminary experiments found about a 7% change rate during voltage sweeps, and the final conclusion reports a 6.3% probability of measurement discrepancies under the stated STS conditions. The system therefore demonstrates sustained autonomous operation, not the elimination of physical instability (Diao et al., 2024).

4. From maintenance to autonomous discovery

Later work broadens AUTOPROBE-like operation from probe maintenance and site selection to discovery-driven experiment design. “Novelty-Driven Target-Space Discovery in Automated Electron and Scanning Probe Microscopy” argues that standard automated microscopy is insufficient when the scientifically important information is not the image itself but the target space of spectra, hysteresis loops, or other functional responses (Pratiush et al., 17 Mar 2026). In that setting, dense mapping is often too slow and may increase beam damage, probe wear, or experimental time costs.

The BEACON framework uses deep kernel learning with a CNN feature extractor and Gaussian process surrogate,

$5.6 \times 5.6$7

and scores novelty in target space through the average distance from a candidate’s predicted response to the $5.6 \times 5.6$8-nearest neighbors in an elite set,

$5.6 \times 5.6$9

Posterior sampling is used for stochastic, uncertainty-aware acquisition. The loop begins with random seed measurements, trains the DKL surrogate, defines an elite set, samples candidate responses, scores novelty, measures the most novel candidate, and updates the dataset. Benchmarking on pre-acquired PFM and STEM-EELS datasets compared BEACON with EI and MU, using MAE of the surrogate, surrogate mean, surrogate uncertainty, target-space coverage, patch-space coverage, and VAE latent-space coverage as diagnostics. In live STEM-EDX, BEACON achieved substantially higher patch-space coverage, while EI and MU tended to cluster spatially (Pratiush et al., 17 Mar 2026).

A closely related but experimentally distinct development appears in “Autonomous microARPES” (Agustsson et al., 2024). Here the problem is not just $^2$0 spatial search but a five-dimensional space $^2$1, with practical scan variables usually $^2$2. The workflow uses asynchronous measurement, dimensionality reduction, and Gaussian-process-based suggestion of the next point. Candidate positions are chosen by maximizing

$^2$3

where $^2$4 is a weighted sum of GP posterior means and variances over scalar tasks such as mean intensity and mean curvature, and $^2$5 penalizes movement time and forbids revisiting prior points. A conventional 2D raster at fixed $^2$6 required about 22,700 points, whereas the autonomous experiment required only a few percent of the data to estimate the task distribution reasonably well; the 3D experiment reached $^2$7 unique positions and showed that a domain low in intensity at $^2$8 or $^2$9 could reveal a well-centered Dirac cone at $28 \times 28$0 (Agustsson et al., 2024).

These systems also delimit what “discovery” means. The BEACON paper explicitly notes the scalarizer bottleneck, and the microARPES paper states that the algorithm’s behavior depends on the selected task, kernel, exploration–exploitation balance, and movement-time penalty. Discovery is therefore guided rather than objective-free (Pratiush et al., 17 Mar 2026, Agustsson et al., 2024).

5. Software mediation, portability, and reusable components

As autonomous probe behavior has expanded, software architecture has become a research topic in its own right. “afspm: A Framework for Manufacturer-Agnostic Automation in Scanning Probe Microscopy” formalizes this layer by defining generic control and data structure schemas passed among independent software components, with final instrument commands produced only after an SPM-specific translator converts the generic request into the microscope’s native API (Sullivan et al., 28 Aug 2025).

afspm focuses on high-level, low-frequency experimental decisions rather than replacing low-level feedback control. Components are instantiated from a TOML configuration, run as independent child processes, and are monitored through heartbeats so that crashed or frozen components can be restarted automatically. Communication uses Google Protocol Buffers for typed serialization and ZeroMQ for in-process, inter-process, or network transport. The framework supports a publisher-subscriber path for microscope data and state changes, and a control-request path for actions. A generic microscope scheduler mediates access and enforces the rule that only one component may control the SPM at a time. It also implements an override mechanism through experiment problem flags: if a non-controlling component detects a problem, such as tip degradation, the scheduler can remove control from the current component and grant it only to a component that claims to resolve the flagged issue (Sullivan et al., 28 Aug 2025).

The paper validates translator portability on two separate systems: GXSM with a softdB Signal Ranger MK2-810 controller, and an Asylum Research MFP-3D-BIO. It also demonstrates a reusable thermal drift correction component based on image registration, keypoint detection, brute-force matching with cross-checking, and RANSAC. Candidate feature methods included SIFT, SURF, BRIEF, ORB, KAZE, and AKAZE; BRIEF was selected because it produced many keypoints and reduced the chance of failure in finding a mapping. In a roughly 16-hour AM-AFM experiment on a CD stamper, the correction tracked a non-linear drift trajectory and largely kept scans aligned (Sullivan et al., 28 Aug 2025).

A plausible implication is that AUTOPROBE has evolved from single-purpose microscope routines into a modular software ecology in which tip classifiers, drift correctors, experiment schedulers, and microscope translators can be recombined across instruments. That implication is consistent with the framework’s stated goal of portability and code reuse (Sullivan et al., 28 Aug 2025).

6. AUTOPROBE as dynamic internal representation selection for code LLMs

In a distinct literature, “AUTOPROBE” names a method for assessing the correctness of LLM-generated code by probing model internals rather than controlling a physical probe (Vu et al., 3 Oct 2025). The problem setting is white-box correctness assessment across compilability, functionality, and security. Prior methods typically probe a fixed layer and token position, such as the last token at the last layer. The paper argues that this is too rigid because the best layer/token combinations vary substantially across DeepSeek Coder, Code Llama, and Magicoder.

The proposed pipeline has three parts: an internal representation extractor, an informative representation selector, and a correctness predictor. Token sampling is boundary-aware, using the first token, last token, first code token, and last code token. Layer sampling is uniform with interval $28 \times 28$1, yielding $28 \times 28$2 sampled layers. On this sampled set $28 \times 28$3, the selector assigns normalized attention scores

$28 \times 28$4

reweights each hidden state as $28 \times 28$5, aggregates the weighted vectors, and feeds the result to a probing classifier. The experimental classifier uses an input layer, two hidden layers of 128 and 64 neurons, and an output layer; aggregation choices include concatenation, summation, mean pooling, max pooling, and min pooling (Vu et al., 3 Oct 2025).

The evaluation spans HumanEval, MBPP, DevEval, SecurityEval, CWEval, CODEGUARD+, and SALLM, with six open-source code LLMs: DeepSeek Coder-1.3B, DeepSeek Coder-6.7B, Code Llama-7B, Code Llama-13B, Magicoder-7B, and Code Gemma-7B. The main reported results are that AUTOPROBE surpasses the state-of-the-art white-box approach by 18% in security assessment and shows robustness to code complexity with performance higher than other approaches by up to 19% for compilability and 111% for functionality. On mixed-language HumanEval generalization, it achieves the best average F1-score of 0.80. The paper also reports that removing attention-based selection drops F1 from 0.76 to 0.69, that boundary-aware sampling gives the best trade-off between performance and efficiency, and that the sampled setting uses about 16.5 GB of memory rather than about 300 GB when all token states are retained (Vu et al., 3 Oct 2025).

This non-microscopy usage is terminologically important because it shows that AUTOPROBE is not confined to hardware instrumentation. In both senses, however, the defining operation is selective probing of internal state in order to make a downstream decision: in one case about whether and how to continue an experiment, and in the other about whether generated code is correct (Vu et al., 3 Oct 2025).