ALD/ALE 2026 Session AF1-WeA: Modeling for ALD Processes I

As atomic layer deposition (ALD) is increasingly applied to complex, multicomponent materials, selecting appropriate compositions and phase windows in practice still relies heavily on trial-and-error. This becomes a serious limitation for functional oxides, where small changes in composition can strongly affect phase formation and electrical properties under thin-film growth conditions. In this work, we present an inverse-design framework for ALD, a general approach that connects recent advances in data-driven artificial intelligence (AI) models for inorganic materials with experimental ALD process development. We first generate many candidate crystal structures by using an AI-based structure generator (a model that proposes plausible crystal structures from composition). This generator prioritizes physically reasonable inorganic structures and compositions, as well as thermodynamically stable candidates (i.e., structures likely to form). These structures are relaxed with machine-learning interatomic potential (a fast surrogate for DFT) to obtain consistent formation energies and the energy above the convex hull (E_hull), enabling fast thermodynamic screening across composition. Based on this thermodynamic screening, graph neural network models are used to estimate key electronic properties, allowing composition–structure–property trends to be mapped without extensive first-principles calculations. The framework is demonstrated using the Hf–Zr–O dielectric system. The model captures a clear trade-off between band gaps and dielectric response across composition and highlights an intermediate composition range where low-energy tetragonal and orthorhombic phases are frequently predicted. Guided by these results, Hf_1-xZr_xO₂ thin films were deposited by atomic layer modulation (ALM; an ALD sequencing method that tunes the Hf/Zr ratio cycle-by-cycle), which enables atomic-scale control of cation ratios within a single ALD sequence. Structural, electrical, and optical measurements show that the predicted phase evolution and property trends are reproduced experimentally, confirming that the model-guided design window remains meaningful under realistic ALD conditions. Overall, this framework provides a practical way to reduce empirical trial-and-error in ALD by focusing experiments on the most promising composition and phase regions. Although demonstrated here using a Hf–Zr–O system, the framework is not material-specific and can be readily applied to other complex oxides, doped systems, and emerging ALD materials.

Atomic layer deposition (ALD) is a key thin-film growth technique in advanced semiconductor manufacturing, enabling atomic-scale thickness control and excellent conformality. Despite the technological maturity of the trimethylaluminum (TMA)–H₂O process for Al₂O₃, establishing a quantitative link between atomistic surface reactions and experimentally observed growth behavior remains challenging due to the complexity of the reaction network, finite-temperature kinetic effects, and steric constraints on precursor adsorption.

In this work, we develop a kinetic Monte Carlo (KMC) framework for the TMA–H₂O ALD process by integrating molecular dynamics (MD) simulations with a neural network potential (NNP) trained to near density-functional-theory accuracy. Compared with our previous study, the present model introduces several key improvements. First, the sticking probability of TMA adsorption, previously approximated as an empirical constant, is re-evaluated using explicit MD collision simulations with a chemically irreversible adsorption criterion, yielding a lower and more realistic value. This update enables more accurate estimation of adsorption timescales and surface saturation behavior. Second, the surface reaction network is substantially expanded to explicitly include hydrogen migration as well as H₂O formation and desorption from hydroxyl groups, which were neglected in earlier modeling. Finally, steric effects arising from the bulky methyl ligands of intact TMA molecules are quantitatively incorporated into the KMC model through a local coordination-based descriptor that restricts adsorption in crowded environments.

The KMC simulations quantitatively reproduce key experimental growth characteristics of Al₂O₃ ALD. The simulated mass evolution within a single ALD cycle captures the characteristic rapid-then-slow mass uptake during the TMA pulse, followed by a negative mass change during the H₂O pulse associated with ligand removal and by-product desorption. The predicted saturated mass gain reaches 36.8 ng cm⁻² cycle⁻¹ at 393 K, in good agreement with in situ quartz-crystal microbalance measurements. Furthermore, the simulations indicate that higher temperature and longer H₂O pulse duration promote more efficient removal of methyl ligands, leading to reduced residual carbon in the growing film. These trends highlight the importance of thermal activation and sufficient reactant exposure for minimizing carbon contamination during Al₂O₃ ALD. Overall, the combined treatment of MD-derived adsorption kinetics, expanded surface reaction pathways, and steric constraints provides a physically grounded and quantitatively improved description of Al₂O₃ ALD growth behavior.

Cu interconnects in advanced ULSI employ a bilayer liner/barrier stack to ensure adhesion to Cu and to suppress Cu diffusion. As scaling progresses, however, the effective resistivity of Cu line increases, motivating the use of a single-layer CoW liner/barrier that can provide both adhesion and barrier functions while improving the Cu volume fraction. A major challenge in forming an ultrathin (1–2 nm) continuous CoW film on dielectric surfaces by ALD is the intrinsically low nucleation density of metal ALD on insulators (<10¹⁰ cm^-2), whereas >10¹⁴ cm^-2 is typically required for immediate film continuity.

To address this nucleation bottleneck, we explore Pd as a catalytic growth enhancer (GE) for CoW ALD. We previously observed that Co ALD on PVD-Pd/SiO₂ increases the Co nucleation density to ~10¹² cm^-2, i.e., ~100× higher than conventional nucleation on dielectrics [1]. Building on this concept, the present work aims to develop a Pd ALD process capable of forming Pd clusters with nucleation densities ≥10¹³ cm^-2 on dielectric surfaces, thereby further enhancing the initial nucleation of CoW.

First, we investigated the chemisorption behavior of the Pd precursor Pd(hfac)₂ using neural-network-potential molecular dynamics (NNP-MD). The NNP was trained to reproduce dispersion-corrected PBE (Perdew–Burke–Ernzerhof) reference data using Matlantis^TM. NNP-MD simulations at 450 K showed no chemisorption of Pd(hfac)₂ on hydroxyl-terminated SiO₂, whereas adsorption was observed on hydroxyl-terminated α-Al₂O₃ within 100 ps. These results suggest that creating an Al₂O₃-like hydroxylated surface is essential to enable high-density Pd nucleation on insulating substrates.

Guided by the simulation, we experimentally performed Pd ALD on SiO₂ and Cu substrates after surface modification with trimethylaluminum (TMA) at 145 °C to form an Al₂O₃-like overlayer. We compared (a) no treatment, (b) TMA treatment followed by air oxidation, and (c) TMA treatment followed by H₂O exposure. On SiO₂, Pd uptake increased after surface modification. On Cu, a marked increase in Pd uptake was obtained only for condition (c). The higher effectiveness of (c) over (b) is attributed to differences in Al₂O₃ surface termination; notably, our MD simulations also indicated that Pd(hfac)₂ does not adsorb on gibbsite-like Al₂O₃ surfaces. These combined computational/experimental results provide a practical route to engineer dielectric surfaces for high-density Pd nucleation as a growth enhancer toward ultrathin, continuous CoW liner/barrier ALD.

Reference

[1] Deng et al., Advanced Metallization Conference 2025, #4-3, Oct. 9-10, 2025, Toyo (2025).

In this work we introduce the design of an atomic layer deposition (ALD) reactor augmented with an AI interface for autonomous materials synthesis. Our modular interface encapsulates the particularities of the hardware behind a Python interface that communicates with the ALD control software via transmission control protocol (TCP). This interface is compatible with model context protocol (MCP) interfaces used in agentic frameworks. To evaluate its performance, we have integrated our tool with a simple AI agent that leverages a large language model to transform user-supplied queries into ALD processes that are then run in our reactor. Our approach uses a JavaScript object notation (JSON) schema to encode ALD processes. Our experimental results show that the AI interface does not impose a significant overhead to our control software, at least within our fastest 10 ms scale. We also carried out a detailed evaluation of the agent performance based on leading models in three types of tasks: basic instruction tasks and process discovery tasks, where the agent is presented with a target material and needs to identify the correct ALD process compatible with the reactor configuration, and process optimization. Despite the simplicity of our agent design, we observed that most of the advanced models excelled at the instruction tasks. However, only recent models such as o1, o3, GPT-5, and Claude Opus 4, with reasoning capabilities, performed well in process discovery tasks. While the results obtained are promising, we identify areas where AI research could help improve the performance of autonomous process discovery and optimization tasks involving atomic layer deposition.

Understanding and ultimately controlling the elementary surface reaction steps that govern atomic layer deposition (ALD) remains a fundamental challenge for rational precursor design, growth selectivity, and process optimization. Although ALD is often idealized as a sequence of perfectly self-limiting half-reactions, practical film growth frequently proceeds through a complex interplay of competing ligand-exchange pathways, surface restructuring, precursor fragmentation, parasitic decomposition, and coverage-dependent kinetics. Quantitative prediction of reaction barriers, mechanistic branching, and rate-determining steps is therefore essential to connect molecular-scale surface chemistry with macroscopic deposition behavior such as growth-per-cycle (GPC), conformity, and selectivity.

Recent advances in high-throughput in silico reaction discovery, automated reaction network construction, and machine-learning interatomic potentials (MLIPs) are enabling mechanistically informed exploration of ALD chemistry at unprecedented length and time scales. By combining algorithmic reaction enumeration, accelerated transition-state refinement, and ML-driven reactive sampling, emerging computational frameworks provide a systematic route to map ALD reaction landscapes, identify kinetic bottlenecks, and evaluate competing pathways across broad precursor–surface chemical spaces.

At Entalpic, we develop agnostic and scalable workflows for the automated construction of complete ALD reaction mechanisms by integrating state-of-the-art cheminformatics and quantum chemistry methodologies. Using bond–electron formalism–based reaction enumeration, we generate comprehensive sets of chemically plausible surface reactions, intermediates, and mechanistic pathways. Thermodynamic screening of reaction networks is rapidly performed using MLIPs, enabling efficient evaluation of large chemical spaces beyond the limits of conventional density functional theory.

To characterize kinetic accessibility, we compute activation barriers through a combination of transition-state sampling strategies, including advanced methods such as Popcornn, alongside generative approaches for transition-state structure proposal and refinement. Through the LeMaterial initiative, we are assembling the largest curated database of transition-metal–organic complex reaction pathways, which directly supports the training of predictive and generative models for mechanistic inference. In parallel, we have benchmarked key elementary reactions for technologically relevant ALD precursors, establishing mechanistic descriptors that correlate with experimentally observed growth-per-cycle trends.