Machine learning and atomic layer deposition: predicting saturation times from reactor growth profiles using artificial neural networks

TL;DR

This work tackles predicting ALD saturation times from thickness growth profiles without requiring explicit surface-kinetics models. It uses a CFD-generated dataset to train feedforward neural networks of varying depth to map thickness vectors and dose time to the saturation time $t_\mathrm{sat}$, revealing that deeper networks achieve high accuracy while remaining robust to limited input data. The study finds that networks with one hidden layer can already closely predict $t_\mathrm{sat}$, whereas overly large two-hidden-layer models may overfit unless carefully regularized; crucially, as few as 8 thickness points can suffice for accurate predictions. The results indicate reactor-specific training data are essential, but a single surrogate model can accelerate ALD process optimization across related conditions, aiding rapid process development and non-ideality detection within the reactor system.

Abstract

In this work we explore the application of deep neural networks to the optimization of atomic layer deposition processes based on thickness values obtained at different points of an ALD reactor. We introduce a dataset designed to train neural networks to predict saturation times based on the dose time and thickness values measured at different points of the reactor for a single experimental condition. We then explore different artificial neural network configurations, including depth (number of hidden layers) and size (number of neurons in each layers) to better understand the size and complexity that neural networks should have to achieve high predictive accuracy. The results obtained show that trained neural networks can accurately predict saturation times without requiring any prior information on the surface kinetics. This provides a viable approach to minimize the number of experiments required to optimize new ALD processes in a known reactor. However, the datasets and training procedure depend on the reactor geometry.

Figures (5)

• Figure 1: a)-c) Shallow and deep neural networks considered in this work d) rectifying linear function (ReLU) used in each of the layers.
• Figure 2: Prediction errors in the testing dataset for one-shot prediction of saturation times from a single growth profile for three different neural networks: A) Shallow network; B) 1 hidden layer ($N=30$); C) 2 hidden layers $N_1=30$, $N_2 = 10$. Data is shown for datasets comprising 20 independent thickness values. The data is color-coded according to how far the dose time used for prediction is to the saturation time (lighter means shorter dose times)
• Figure 3: Prediction accuracy as a function of the number of independent points in the growth profile. Top: mean relative error. Bottom: standard deviation. Results are shown for the shallow network and deep networks with 1 and 2 hidden layers.
• Figure 4: Saturation time prediction accuracy using a 1 layer deep neural network ($M=30$) for different numbers of data points in the growth profile: A) $N=4$, B) $N=5$ and C) $N=10$. The data are color-coded according to how far the dose time used for prediction is from the saturation time (lighter means shorter dose times)
• Figure 5: Prediction accuracy for a network with one hidden layer as a function of the number of independent points in the growth profile for different numbers of neurons, $M$, in the hidden layer.