Towards a virtual ALD reactor: Computational chemistry and machine learning in fruitful interaction
Prof. Dr. Ralf Tonner-Zech (Universität Leipzig)
Transistor downscaling and the miniaturization of key components have been the primary drivers of performance improvements in microelectronic devices for decades. We are currently entering the “Angstrom era,” where the most critical device features are only a few atoms wide. This necessitates unprecedented atomic precision during structure fabrication, a task becoming increasingly challenging for the dominant top-down lithography approaches. Consequently, bottom-up approaches are gaining significance for specific industrial manufacturing steps. Atomic Layer Deposition (ALD) is a key enabling technology here: it facilitates the deposition of defined, conformal thin films on high-aspect- ratio structures, which is essential for modern memory devices, for example. This is achieved through the self-limiting surface reactivity of a precursor and a co-reactant. The need for atomic control requires a deep understanding of the chemical reactivity between the substrate and the precursor. Furthermore, Area-Selective ALD (AS-ALD) is becoming vital, where materials are deposited selectively on one substrate while leaving neighboring surfaces bare. The prevalent inhibitor-based AS-ALD approach necessitates understanding the reactivity of inhibitor molecules with the substrate surface as well. The experimental screening and optimization of these surface reactions are highly resource intensive. For every substrate-precursor combination, process parameters must be individually optimized, and the resulting film quality must be analyzed using demanding metrology techniques.
Computational approaches promise a significant acceleration of this material discovery process by identifying the most promising substrates and precursors. The main challenge lies in accurately modeling the complex chemistry and physics characteristic of typical ALD reactions. This demands large system sizes, extended simulation timescales, and simultaneously maintaining high accuracy.
In this talk, I will present our ongoing efforts towards developing a “virtual ALD reactor” capable of enabling this computational discovery approach.
- Ab initio methods, such as Density Functional Theory (DFT), help to elucidate key reaction mechanisms at the atomic level and obtain information on surface coverage [1].
- Machine Learning (ML), combined with Molecular Dynamics (MD) simulations, allows us to cover much larger timescales to model the dynamic behavior of systems [2].
Supplemented by adsorption modeling using Random Sequential Adsorption (RSA) or Kinetic Monte Carlo (kMC), these integrated approaches can yield significant computational insights into a variety of material systems [3].
The ultimate goal is to establish an efficient and precise predictive framework for both ALD and AS- ALD processes.
[1] F. Pieck, R. Tonner-Zech, Chem. Mat., 2025, 37, 2979.
[2] H. Weiske, R. Barrett, R. Tonner-Zech, P. Melix, J. Westermayr, 2025, arXiv: 2509.14828.
[3] (a) J. Yarbrough, F. Pieck, A. B. Shearer, P. Maue, R. Tonner-Zech, S. F. Bent, Chem. Mater. 2023, 35, 5963; (b) J. Yarbrough, F. Pieck, D. Grigjanis, I.-K. Oh, P. Maue, R. Tonner-Zech, S. F. Bent, Chem. Mater. 2022, 34, 4646; (c) S. Zoha, F. Pieck, B. Gu, R. Tonner-Zech, H.-B.-R. Lee, Chem. Mater. 2024, 36, 2661; (d) P. P. Wellmann, F. Pieck, R. Tonner-Zech, Chem. Mater. 2024, 36, 7343.
