🔬 materials science

Multitask Scanning Probe Microscopy

This paper introduces "multitask scanning probe microscopy," an autonomous, closed-loop workflow that utilizes a multitask Gaussian process to dynamically select both measurement locations and experimental protocols, thereby enabling efficient, large-scale nanoscale characterization by intelligently balancing rapid imaging with slower, more invasive modalities.

Original authors: Aditya Raghavan, Yu Liu, Ian Mercer, JP Maria, Sergei Kalinin

Published 2026-08-11

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Aditya Raghavan, Yu Liu, Ian Mercer, JP Maria, Sergei Kalinin

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Detective's Dilemma: Mapping the Invisible World

Imagine you are a detective trying to solve a mystery on a giant, invisible city block. You have a special magnifying glass that can see tiny details, but it comes in two flavors: one is a quick, gentle peek that tells you about the shape of the streets, and the other is a slow, heavy-handed investigation that reveals the electrical wiring and secrets of the buildings, but it takes forever and might accidentally knock a few things over. In the world of materials science, this "city block" is a microscopic surface of a material, and the "magnifying glass" is a tool called a Scanning Probe Microscope (SPM). Scientists use these tools to understand how materials behave, which is crucial for building better electronics, batteries, and solar cells.

The big problem scientists face is that these materials are huge compared to the tiny details they need to see. A single computer chip or a new type of metal sheet might have thousands of spots to check. If they try to use the slow, heavy-handed investigation on every single spot, it would take years and might ruin the sample. If they only use the quick peek, they miss the important secrets. For a long time, the solution was to just do both at every spot, wasting time and energy. But what if the microscope could be smart enough to decide where to look and which tool to use, learning from one spot to guess what's happening at the next? That is the challenge this paper tackles: teaching a microscope to be an autonomous detective that knows when to be gentle and when to dig deep, saving time while still solving the mystery.

The Paper's Story: A Smart Microscope That Learns to Choose

In this study, the researchers introduced a new way of using scanning probe microscopy called "multitask scanning probe microscopy." Think of it as giving the microscope a brain that can juggle two different jobs at once. Usually, when scientists map a material, they have to decide: "Do I scan this whole area with the fast, gentle mode, or do I stop at every single point to do the slow, detailed test?" Doing both everywhere is too slow, and doing just one leaves gaps in knowledge.

The team created a system where the microscope doesn't just pick a spot to look at; it also picks how to look. They used a mathematical trick called a "multitask Gaussian process." Imagine you are trying to guess the weather in a whole city. You have two sensors: one that measures temperature (fast and easy) and one that measures humidity (slow and tricky). If you know that temperature and humidity usually go hand-in-hand in this city, you don't need to measure humidity everywhere. If you measure the temperature in a few new spots, your smart brain can guess the humidity in those spots based on what it learned from the temperature.

In the lab, the researchers tested this on a special wafer (a thin slice of material) made of Aluminum Scandium Nitride. They set up a grid with over 1,100 possible spots to measure. They defined two "tasks":

Task A: Tapping-mode imaging, which is like a gentle tap-tap-tap to feel the roughness of the surface. It's fast and doesn't hurt the sample.
Task B: DART-mode imaging, which is a more complex, contact-based measurement that also feels roughness but interacts with the surface differently and takes longer.

The experiment started with a "learning phase." The microscope went to five random spots and did both the gentle tap and the complex contact test. This taught the system how the two measurements were related. They found a strong connection: when the surface was rough in the gentle tap, it was usually rough in the complex test too, with a correlation score of 0.752.

Once the system learned this relationship, the real magic happened. The microscope entered a "closed-loop" mode, meaning it could think, decide, and act all by itself. It looked at the map of what it knew and asked, "Where do I need to look next, and which tool should I use to learn the most?"

Instead of checking every spot with both tools, the microscope started picking one spot and doing only one of the two tests. If it did the gentle tap at a new spot, the system used the learned relationship to update its guess about the complex test at that same spot, and vice versa. Over 25 steps, the microscope made 25 new measurements. It chose to do the complex DART test 14 times and the gentle tapping test 11 times, based entirely on where it thought it would learn the most.

The result was impressive. By the end of the experiment, the microscope had mapped the roughness of the entire wafer with high confidence, even though it only performed one type of measurement at each location. It managed to reduce the uncertainty in its maps by about 80% for both tasks. The system proved that you don't need to measure everything twice to understand everything; you just need to be smart about how you mix your measurements.

The paper explicitly rules out the idea that you must perform a predefined sequence of measurements or that you need to measure every single spot with every single tool. They showed that a fixed, rigid plan is inefficient. However, they also noted that this specific success relied on the two tasks being strongly related (both measuring roughness). They suggest that for tasks that are very different or weakly related, the system might need more complex math to work, but the foundation is now there.

In short, the researchers built a microscope that acts like a curious, efficient explorer. It learned that a quick glance at the surface could tell it a lot about the deeper properties, allowing it to skip unnecessary steps. This approach doesn't just save time; it opens the door to exploring massive, complex materials libraries that were previously too big or too delicate to map in full detail. The authors suggest that this method could eventually be used to mix and match even more different types of measurements, like combining surface shape with electrical current or magnetic fields, guiding the microscope to the most important spots to study next.

Technical Summary: Multitask Scanning Probe Microscopy

Problem Statement
Scanning probe microscopy (SPM) offers nanoscale access to diverse material properties (structural, electrical, electromechanical, magnetic, and mechanical). However, applying SPM to large-area samples, such as semiconductor wafers or combinatorial material libraries, presents a significant efficiency challenge. The number of potential measurement locations can reach thousands, while different measurement modalities vary drastically in acquisition time and the potential for tip or sample damage. Exhaustive multimodal mapping across spatial grids is often impractical. While previous autonomous SPM workflows have successfully optimized where to measure using active learning, they typically assume the measurement protocol (modality) is fixed beforehand. This limitation prevents the autonomous allocation of measurement resources between fast, weakly perturbative imaging and slower, potentially damaging spectroscopic or contact-based measurements.

Methodology
The authors introduce Multitask Scanning Probe Microscopy, a live, closed-loop framework that autonomously selects both the next measurement location and the specific experimental protocol (task).

Statistical Framework: The approach utilizes a Multitask Gaussian Process (GP) based on an Intrinsic Coregionalization Model (ICM). In this model, different SPM modalities (e.g., tapping-mode topography vs. DART imaging) are treated as related prediction tasks sharing a common underlying material structure but possessing distinct spatial response landscapes and noise characteristics.
Model Architecture: The covariance between any two location-task pairs is defined as the product of a spatial kernel ( $K_x$ ) and a task-covariance matrix ( $B$ ). The spatial kernel uses a Matérn 5/2 function with automatic relevance determination to accommodate sharp spatial variations. The task-covariance matrix captures correlations between modalities, allowing a measurement in one mode to update predictions for other modes across the entire domain.
Active Learning Loop:
1. Initialization: Paired measurements (both tasks) are acquired at a few "seed" locations to establish the initial cross-task correlation.
2. Selection: At each iteration, the controller evaluates the joint posterior distribution. It employs a random-scalarization Upper Confidence Bound (UCB) acquisition function. Task weights are sampled from a symmetric Dirichlet distribution to stochastically balance exploration between tasks.
3. Execution: The microscope autonomously configures the specific mode (e.g., switching between tapping and contact-resonance), executes the measurement, and extracts a scalar descriptor (e.g., surface roughness).
4. Update: The joint model is updated, refining predictions for both the measured task and the unmeasured tasks via the learned correlations.
5. Exploration: To prevent spatial clustering, forced exploration steps are inserted periodically, selecting the location farthest from previous measurements.

Experimental Implementation
The framework was implemented on an automated Oxford Instruments Asylum Research Jupiter AFM with a large-sample stage. The demonstration utilized a composition-spread AlScN wafer with a candidate grid of 1,117 positions.

Tasks: Two tasks were defined: (1) Tapping-mode roughness and (2) Dual AC Resonance Tracking (DART) roughness.
Protocol: The experiment began with 5 paired seed locations (10 total observations). This was followed by 25 active-learning steps where only one task was measured per location.
Total: 35 individual scans were performed to reconstruct the roughness landscapes for both modalities.

Key Results

Correlation Learning: The model successfully learned a strong positive correlation ( $\rho = 0.752$ ) between tapping-mode and DART roughness, despite the tasks having different physical interactions and noise levels.
Uncertainty Reduction: The joint model reduced the prior standard deviation of the roughness landscapes by approximately 80% for both tasks.
- Tapping-mode posterior uncertainty: Reduced from 0.925 nm to 0.176 nm.
- DART-mode posterior uncertainty: Reduced from 1.098 nm to 0.235 nm.
Efficiency: The system achieved this reconstruction accuracy by measuring only one modality at each location after the seed phase, effectively distributing the measurement budget across spatial and modality dimensions without acquiring redundant data.
Noise Characterization: The model correctly identified that DART measurements were approximately twice as noisy as tapping-mode measurements (0.437 nm vs. 0.233 nm).

Significance and Claims
The paper claims to demonstrate the first live, closed-loop autonomous selection of both measurement location and modality on an operating microscope. The primary significance lies in extending active learning in SPM from purely spatial sampling to the autonomous allocation of measurement modalities.

The authors state that this framework provides a basis for combining rapid, weakly perturbative imaging with slower, contact-based, or state-changing measurements. By leveraging statistical correlations between tasks, the system can construct accurate response landscapes for multiple properties without performing exhaustive multimodal mapping at every point. The authors position this as a foundational step toward more complex applications, such as combining topography with electrical or electromechanical measurements, where cost and sample damage are critical constraints. The work establishes the instrumental and computational basis for multitask SPM, with the code and workflow made openly available.

The Detective's Dilemma: Mapping the Invisible World

The Paper's Story: A Smart Microscope That Learns to Choose

More like this