🔬 materials science

Active learning molecular beam epitaxy of complex quantum materials

This paper presents a data-efficient active learning framework combining a random forest surrogate model with sequential model-based optimization to autonomously navigate the complex, non-linear growth landscape of molecular beam epitaxy, successfully identifying optimal synthesis conditions for the metastable topological Weyl ferromagnet Fe $_3$ Sn within just a few iterations.

Original authors: Raghutheja Bollampally, Soumya Sankar, Yuqi Qin, Berthold Jäck

Published 2026-08-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Raghutheja Bollampally, Soumya Sankar, Yuqi Qin, Berthold Jäck

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

Making new materials often feels like searching for a needle in a haystack, but the haystack is made of invisible, shifting clouds of temperature and chemical mix. Scientists who build thin films of matter, layer by layer, face a daunting task: they must navigate a vast landscape of possibilities to find the single, perfect combination of conditions that creates a high-quality material. For decades, this search has relied on human intuition, where researchers tweak one setting at a time, hoping to stumble upon the right spot. This trial-and-error approach is slow, expensive, and often fails when the material is complex, because the perfect conditions might be a tiny, sharp peak surrounded by a sea of failure. Now, a team of researchers has introduced a smarter way to explore this landscape, using a computer program that learns as it goes, turning a years-long search into a matter of days.

The material at the center of this story is a specific type of metal compound called Fe3Sn, which holds promise for advanced magnetic sensors and future computing technologies. To create this material, scientists use a technique called molecular beam epitaxy, where they heat up pure elements in a vacuum chamber and let them drift onto a hot surface to form a crystal. The process is incredibly sensitive; if the surface is too hot, the atoms fly off, and if it is too cold, they clump together incorrectly. Furthermore, the ratio of iron to tin atoms must be just right, or the material will form a different, useless substance. In the past, finding the right settings for such a finicky material required dozens of experiments, each one taking hours to set up and analyze. The researchers wanted to know if a machine could learn to find the perfect settings faster than a human could, but they faced a problem: the standard computer tools used for this job assume that the world changes smoothly, like a gentle hill. However, the world of these complex materials is full of sudden cliffs and sharp edges, where a tiny change in temperature causes the material to fail completely.

To solve this, the researchers built a new kind of learning system that does not assume the landscape is smooth. They started with a small collection of data from about seventeen previous experiments, which gave them a rough map of where the material worked and where it did not. Instead of using a traditional smoothing algorithm, they employed a method that acts like a forest of decision trees, a technique that is excellent at spotting sharp boundaries and sudden changes. This computer model looked at the data and realized that the most important factor was not the exact mix of atoms, but the temperature of the surface and the quality of the starting material. The model predicted that there was a specific "plateau" of success—a range of temperatures and atom ratios where the material would grow perfectly—surrounded by areas where it would fail.

The team then let the computer take the wheel. In a closed loop, the machine chose the next set of conditions to test, the robot grew the film, and then the film was analyzed to see how well it turned out. The results were fed back into the computer, which updated its map and chose the next best experiment. Remarkably, the system did not just guess; it learned. After just four new experiments guided by the computer, the accuracy of its predictions doubled, narrowing the error down to about ten percent. The machine successfully identified the narrow window where the material grows well, a task that would have taken a human researcher many more attempts to find. The computer also revealed a surprising truth: once the temperature was set correctly, the material was surprisingly forgiving of small changes in the atom mix, a detail that human intuition might have missed.

This work demonstrates that machines can now navigate the treacherous, high-stakes terrain of complex material synthesis without human guidance. By replacing old, smooth-sounding models with a system that understands sharp edges and sudden shifts, the researchers have shown that autonomous discovery is possible even for the most difficult quantum materials. The result is a path toward laboratories where robots can design, build, and test new materials on their own, freeing scientists to focus on the big questions of what these materials can do, rather than the tedious work of finding the right settings to make them.

Technical Summary: Active Learning Molecular Beam Epitaxy of Complex Quantum Materials

Problem Statement
The synthesis of complex quantum materials via Molecular Beam Epitaxy (MBE) requires navigating high-dimensional, non-equilibrium thermodynamic phase spaces. Unlike bulk synthesis, MBE is governed by delicate control parameters (e.g., substrate temperature, stoichiometry) that determine surface kinetics and chemistry. Conventional optimization relies on manual, one-at-a-time parameter sweeps, which are time-consuming and heavily dependent on researcher intuition. While Machine Learning (ML) and Bayesian Optimization (BO) have been applied to thin-film growth, existing approaches predominantly utilize continuous surrogate models (e.g., Gaussian Process Regression) that assume smooth parameter landscapes. These models struggle to capture the abrupt crystallographic phase boundaries and narrow growth windows inherent to binary quantum materials, such as metastable line compounds. Consequently, it remains an open question whether active learning can effectively navigate the constrained, high-dimensional phase spaces required for complex binary compounds where competing secondary phases make manual optimization difficult.

Methodology
The authors propose an Active Learning (AL) protocol based on Sequential Model-Based Optimization (SMBO) designed for the closed-loop MBE of the topological Weyl ferromagnet Fe $_3$ Sn. The framework addresses the limitations of standard continuous models through the following components:

Surrogate Model: Instead of Gaussian Process Regression (GPR), the authors employ a Random Forest (RF) surrogate model. RF was selected because its hierarchical space-partitioning architecture naturally captures highly non-linear relationships, sharp phase transitions, and discontinuities without assuming a smooth underlying topology. This is critical for modeling the abrupt thermodynamic boundaries of Fe $_3$ Sn.
Acquisition Function: The system utilizes an Expected Improvement (EI) function to balance exploration and exploitation. The RF model provides both a predictive mean (exploitation) and a measure of predictive uncertainty (exploration) derived from the variance across decision trees.
Feature Engineering and Scoring:
- Inputs: Fe-to-Sn beam flux ratio, substrate heater power (used as a proxy for temperature to avoid calibration errors), and pre-growth Pt(111) buffer layer quality (extracted via in-situ RHEED).
- Targets: A "Holistic Additive Score" ( $S_{Total}$ ) ranging from 0–100, calculated as a weighted sum of three automated metrics: RHEED surface morphology (50% weight), EDS stoichiometry (25%), and XRD bulk crystallinity (25%).
- Automation: Automated algorithms extract scalar metrics from XRD (FWHM of the Fe $_3$ Sn(0002) peak) and RHEED (horizontal FWHM for in-plane crystallinity, vertical intensity variance for 3D islanding, and central streak width for mosaicity).
Workflow: The loop initializes with a sparse dataset of 17 historical growth iterations. The SMBO framework then iteratively trains the RF model, maximizes the EI function to propose the next experimental condition, executes the growth, and feeds the new characterization data back to retrain the model.

Key Results

Model Performance: When compared to GPR on the sparse 17-iteration dataset, the RF model demonstrated superior ability to map the underlying thermodynamic topology. While GPR exhibited "kernel collapse," smoothing over abrupt boundaries and defaulting to the global mean, the RF model accurately captured the non-linear growth landscape and sharp phase transitions.
Optimization Landscape: The SMBO framework identified a robust "Plateau of Success" centered around a filament power of ~3.6 W and a Fe:Sn flux ratio of ~0.8. This plateau spans a temperature range of approximately 30 K, indicating that high-quality Fe $_3$ Sn films are robust to minor fluctuations within this window.
Feature Importance: The RF model autonomously identified substrate filament power (temperature) and initial substrate quality as the dominant factors influencing film quality (importance scores of ~~0.41 and ~0.37, respectively), outweighing the Fe/Sn flux ratio (~~0.22). This suggests that precise temperature control and substrate preparation are more critical than fine-tuning flux ratios once the optimal window is established.
Convergence: Starting from the initial 17 iterations, the active learning loop required only four additional iterations (Runs 18–21) to converge. The absolute predictive error was halved from an initial spike (Run 18, >20.0) to approximately 10% (Run 21, 9.9). The initial error spike was attributed to the EI function deliberately probing high-uncertainty regions to map uncharted boundaries, a necessary step for global optimization.

Significance and Claims
The paper claims to demonstrate a data-efficient framework that enables the autonomous discovery and synthesis of functional quantum materials. By replacing standard continuous surrogate models with a Random Forest architecture, the authors successfully navigated the complex, constrained thermodynamic phase space of a metastable line compound (Fe $_3$ Sn) in only 21 total iterations. This contrasts with previous BO applications on complex materials (e.g., perovskite oxides) that often required 24–44 iterations.

The authors position this work as a critical step toward fully autonomous MBE laboratories. They argue that the RF-based SMBO approach overcomes the "smoothness assumption" limitation of traditional BO, making it suitable for materials with narrow growth windows and competing phases. While the current protocol automates the optimization and decision-making process, the authors note that the next logical step involves physical integration with automated effusion cell control and robotic sample handling to create "self-driving" MBE laboratories. The study does not claim to have solved all challenges in autonomous synthesis but establishes a robust methodological foundation for reducing human intervention in the optimization of complex quantum materials.

More like this