Efficient characterization of surface compositions across high-dimensional materials spaces is critical for accelerating the discovery of surface-dominated functional materials. While X-ray photoelectron spectroscopy (XPS) allows detailed surface composition investigation, it remains a time-intensive technique [1]. For thin-film materials libraries with continuous gradients with hundreds of well-defined compositions, characterizing entire libraries with XPS is considered impractical, as measuring a single surface composition can take up to 1-2 hours.

In order to decrease characterization times significantly, active learning can be used to iteratively select the next measurement area by building and updating a machine learning model during the measurement procedure [2-4]. In this work, we apply a multi-output Gaussian process in an active learning setting, predicting the surface compositions from rapidly acquired volume composition data obtained by energy-dispersive X-ray spectroscopy (EDX). We show how the model’s predictions can still provide a substantial efficiency increase in case it cannot be implemented directly into the measurement device through an application programming interface (API).

To reliably evaluate the prediction accuracy, the effort was made to quantify the surface composition of an entire exemplary materials library in the system Mg-Mn-Al-O [5], resulting in a total XPS measurement duration of 12 complete days. Our results demonstrate that only 13 evenly distributed measurements are sufficient to predict the surface composition of the entire library with an accuracy of 96%, reducing the total measurement time to just 11 hours [6]. This method offers a robust, scalable, and data-efficient approach for integrating advanced surface characterization into materials discovery workflows.

References:

[1] D.R. Baer, K. Artyushkova, C. Richard Brundle, J.E. Castle, M.H. Engelhard, K.J. Gaskell, J.T. Grant, R.T. Haasch, M.R. Linford, C.J. Powell, A.G. Shard, P.M.A. Sherwood, V.S. Smentkowski, Practical guides for x-ray photoelectron spectroscopy: First steps in planning, conducting, and reporting XPS measurements, Journal of Vacuum Science & Technology A: Vacuum, Surfaces, and Films 37 (2019). https://doi.org/10.1116/1.5065501.

[2] A.G. Kusne, H. Yu, C. Wu, H. Zhang, J. Hattrick-Simpers, B. DeCost, S. Sarker, C. Oses, C. Toher, S. Curtarolo, A. V. Davydov, R. Agarwal, L.A. Bendersky, M. Li, A. Mehta, I. Takeuchi, On-the-fly closed-loop materials discovery via Bayesian active learning, Nature Communincations 11 (2020) 5966. https://doi.org/10.1038/s41467-020-19597-w.

[3] T. Lookman, P. V Balachandran, D. Xue, R. Yuan, Active learning in materials science with emphasis on adaptive sampling using uncertainties for targeted design, NPJ Computational Materials 5 (2019) 21. https://doi.org/10.1038/s41524-019-0153-8.

[4] F. Thelen, L. Banko, R. Zehl, S. Baha, A. Ludwig, Speeding up high-throughput characterization of materials libraries by active learning: autonomous electrical resistance measurements, Digital Discovery 2 (2023) 1612–1619. https://doi.org/10.1039/D3DD00125C.

[5] F. Lourens, E. Suhr, A. Schnickmann, T. Schirmer, A. Ludwig, High‐Throughput Study of the Phase Constitution of the Thin Film System Mg–Mn–Al–O, Advanced Engineering Materials 26 (2024). https://doi.org/10.1002/adem.202302091.

[6] F. Thelen, F. Lourens, A. Ludwig, Accelerating Surface Composition Characterization of Thin-Film Materials Libraries using Multi-Output Gaussian Process Regression, arXiv:2503.23471 (2025) https://doi.org/10.48550/arXiv.2503.23471.