On Simulating Thin-Film Processes at the Atomic Scale Using Machine Learned Force Fields

TL;DR

The paper addresses the challenge of realistically simulating thin-film processes at the atomic scale by developing machine-learned force fields (MLFFs) based on Moment Tensor Potentials (MTP). It introduces a data-efficient training pipeline that combines structure-generation protocols with active learning to build MTPs capable of representing gas–surface chemistry across molecule, bulk, surface, and interface domains, integrated with Surface Process Simulation in QuantumATK. Two technologically relevant case studies—precursor pulse in HfO2 atomic layer deposition and ALE of MoS2—demonstrate accurate replication of DFT reference data (RMSEs on the order of tens of meV per atom for energies and a few tenths of eV per Å for forces) and reveal quantitative insights into adsorption probabilities, self-limiting behavior, and chlorine-enhanced etching mechanisms. The results highlight the practical potential of the MTP–SPS framework to accelerate process understanding and optimization for advanced materials fabrication.

Abstract

Atomistic modeling of thin-film processes provides an avenue not only for discovering key chemical mechanisms of the processes but also to extract quantitative metrics on the events and reactions taking place at the gas-surface interface. Molecular dynamics (MD) is a powerful computational method to study the evolution of a process at the atomic scale, but studies of industrially relevant processes usually require suitable force fields, which are in general not available for all processes of interest. However, machine learned force fields (MLFF) are conquering the field of computational materials and surface science. In this paper, we demonstrate how to efficiently build MLFFs suitable for process simulations and provide two examples for technologically relevant processes: precursor pulse in the atomic layer deposition of HfO2 and atomic layer etching of MoS2.

Figures (18)

• Figure 1: Schematic for batch learning.
• Figure 2: Schematic for active learning.
• Figure 3: Schematic showing different training data domains needed to be sampled for training data generation. Four data domains are identified, namely molecule domain, bulk domain, surface domain and molecule-surface interface domain.
• Figure 4: Data flow in the MTP training procedure for ALD and ALE process simulation. Eight step procedure is laid out in 5 levels. The steps are split into levels to enhance the efficiency of training data generation. Steps in each level can be executed in parallel as they are independent of each other. First level has three steps where the three different domains, namely molecules, bulk and surface are sampled independently. The data generated in first level is passed on to the second level. In the second level there are two steps, namely surface active learning and adsorbed surface generation, which can be run in parallel. The third, fourth and fifth levels have just one step each since they are dependent on the data from their preceding level. Step-6 in third level uses all the data generated so far in previous levels to perform active learning of adsorbed surfaces. Step-7 is used to perform active learning of gas surface collisions. Step-8 is optional and can be used to improve the training data for specific process conditions of interest.
• Figure 5: SPS flowchart showing the simulation procedure in panel (a). System setup for MD in SPS is shown in panel (b).