ALD/ALE 2022 Session AF2-MoA: Simulation, Modeling, and Theory of ALD I

In particular, we investigated the relevance of such simulations for the assessment of synthetic pathways and reactions occurring during Atomic Layer Deposition processes. We also assessed the interest of assisting analytical efforts with molecular simulations, for example for substance identification via spectra simulation.

Our work showed that molecular simulations are a useful tool for the development of new ALD precursors. They can be used to different degrees of accuracy to understand:

• Intrinsic molecule stability, either isolated or in a condensed phase
• Reaction thermodynamics and reaction paths
• Surface adsorption/desorption and reactivity

However the predictive power remains to be improved, and at this point, it should be seen as a useful companion to experimental work. Effective collaboration between the synthetic chemists, ALD/CVD practitioners and molecular modeling theorists accelerates the acquisition of a chemical and mechanistic comprehension of ALD Precursors. Such learning can bring an acceleration to the design and development process.

In atomic level processing, the most important requirement is reactivity of a precursor towards deposition or etching of the target material; however, it is difficult to quantify reactivity. One possible experimental measure is the growth per cycle (ALD) or etch per cycle (ALE) under conditions where the chemistry is self-limiting. Experimentally measuring reactivity for all possible precursors and target materials is prohibitively expensive, necessitating alternative ways to explore the large design space.In this talk, we examine how to compute precursor reactivity with atomistic modeling.

We discuss long-standing models [1] that use density functional theory (DFT) to simulate ligand exchange in the precursor molecule itself as a measure of reactivity. Advantages of these models are conceptual simplicity, ease of computation and suitability for processing large sets of molecules [2]. Furthermore, these models can extend from oxides/nitrides to metals, approximate temperature effects and deliver formation energies.

However, more information about the multistep reaction mechanism is required to calculate the growth/etch rate of the process.We examine the thermodynamic basis that dictates how many reaction steps the precursor can undergo at the surface, i.e. the ‘extent’ of the reaction [3].The reaction extent can be converted into a growth/etch rate per cycle if additional information about ligand size and material density is calculated.We propose a high-throughput approach for computing surface reactions with DFT and hence quantifying the reaction extent and the growth per cycle.We then validate our approach against experimental data for commonly used precursors for the ALD of dielectrics.

Finally, we explore the capabilities of machine learning models (ML) to accurately predict the DFT reactivities using structure-property relationships. Since the computational time for running machine learning models is significantly lower than for DFT, an ML approach can enable thousands of ligands to be screened quickly.

These automated modeling approaches enable the quantification of precursor reactivity in terms of growth/etch rates, which will open new opportunities for in silico design and process understanding.

[1] P. Hirva et al., Surf. Sci. 220, 137–51 (1989); S. D. Elliott, Surf. Coat. Technol. 201, 9076–81 (2007); T. P. Holme et al., J. Phys. Chem. A 111, 8147–51 (2007); I. Kazadojev et al., Chem. Vap. Depos. 19, 117–124 (2013).

[2] RASIRC® ALD Oxide/Nitride Wizards™ https://www.rasirc.com/ald-oxide-wizard/ https://www.rasirc.com/ald-nitride-wizard/

[3] S. D. Elliott, Comput. Mater. Sci. 33, 20–5 (2005).

Atomic layer deposition (ALD) processes are known to deposit a sub-monolayer of material per cycle due to (i) steric hindrance effects and/or (ii) limited availability of surface sites.[1, 2] Another contributing factor, which has not been addressed in detail, is that molecules arrive one by one at the surface and adsorb at random surface sites, such that the saturation coverage of precursor ligands substantially deviates from closed packing. In this work, the effect of the random sequential adsorption (RSA) mechanism on the saturation of precursor adsorption and hence on the growth per cycle (GPC) of ALD processes is investigated. By systematically varying the precursor radius and the precursor binding sites in the RSA simulations, the surface densities of the adsorbed precursor molecules are modelled. It is found that the random sequential adsorption of the precursor molecules results in a 25% to 68% lower surface density than the (unrealistic) reference case of closed packing of ligands. Moreover, RSA simulations are also used to model the growth-per-cycle values based on the precursor shape and size. Using this RSA model for metallocene precursor adsorption, GPC values of Co, Os, and Ru ALD processes are calculated as 0.28, 0.38 and 0.37 Å/cycle, in agreement with the experimental findings with accuracy of ± 0.08 Å/cycle. [3-5] Therefore, this work clearly demonstrates the importance of the random sequential adsorption mechanism on the ALD growth.

[1] R.L. Puurunen, Chem. Vap. Depos. 9, 249 (2003).

[2] R.L. Puurunen, Appl. Surf. Sci. 245, 6 (2005).

[3] J. Hämäläinen, T. Sajavaara, E. Puukilainen, M. Ritala, and M. Leskelä, Chem. Mater. 24, 55 (2012).

[4] T. Aaltonen, A. Rahtu, M. Ritala, and M. Leskelä, Electrochem. Solid-State Lett. 6, (2003).

Atomic layer deposition (ALD) of materials has been a key enabling technology for atomically fine-tuned manufacturing of semiconductor devices in the past two decades. As the future brings continued down-scaling of electronic devices, extreme fabrication limits require optimization of existing and development of new deposition setups for maximum process capability and yield, enabled by physical modeling with topography tools, such as Sentaurus Topography.

In this talk, we will show how machine learning (ML) based atomistic simulations of ALD processes can efficiently provide important parameters for topography simulations, which have been historically extracted from experiments. This multiscale modeling approach can be of particular importance when developing new deposition processes where experimental data are unavailable or expensive and time-consuming to obtain. We will demonstrate how to efficiently obtain ALD parameters such as sticking coefficient and maximum precursor coverage by simulating the HfCl₄ precursor pulse in the ALD of HfO₂.

ALD atomistic simulations can be computationally very demanding or even unmanageable if performed using ab initio density functional theory (DFT) molecular dynamics (MD), whereas computationally cheaper conventional Force Fields are often not available and require a long and error-prone process to be developed, as in the case for HfCl₄-HfO₂. This can be overcome by generating and using ML based interatomic potentials (MLIPs). In this work, we efficiently developed and used moment tensor potentials (MTPs), which provide high accuracy with lower computational cost when compared to other MLIPs [1,2]. MTP was trained using reference data from DFT simulations to reproduce ab initio quality potential energy surface of the HfCl₄-HfO₂ system. MD simulations with such trained MTPs give us DFT quality MD trajectories at a significantly (100x) reduced computational effort. Automated MTP training and MD simulations were performed using the QuantumATK software [3,4] developed by Synopsys.

MD simulations of HfCl₄ molecules impacting the HfO₂ (111) substrate reveal that Hf atoms bind to O atoms on the surface as expected, with binding energies being consistent with DFT results. Finally, we show that the sticking coefficient of the precursor decreases as its surface coverage increases and becomes 0 after attaining a maximum coverage at 0.3 monolayers.

[1] A. V. Shapeev, Mult. Model. Sim., 14, 1153 (2016).
[2] Y. Zuo et al., J. Phys. Chem. A, 124, 731 (2020).
[3] S. Smidstrup et al. , J. Phys.: Conden. Matter 32, 015901 (2019).
[4] QuantumATK version T-2022.03, Synopsys QuantumATK (https://www.synopsys.com/silicon/quantumatk.html)