AVS 66th International Symposium & Exhibition
    Thin Films Division Friday Sessions
       Session TF-FrM

Paper TF-FrM4
Process Optimization in Atomic Layer Deposition Using Machine Learning

Friday, October 25, 2019, 9:20 am, Room A216

Session: Theory and Characterization of Thin Film Properties
Presenter: Noah Paulson, Argonne National Laboratory
Authors: A. Yanguas-Gil, Argonne National Laboratory
S. Letourneau, Argonne National Laboratory
A.U. Mane, Argonne National Laboratory
N.H. Paulson, Argonne National Laboratory
A.N. Lancaster, Argonne National Laboratory
J.W. Elam, Argonne National Laboratory
Correspondent: Click to Email
Process development and process optimization are ubiquitous, resource-intensive tasks in thin film research and development. The goal of these activities is to find the set of process parameters (e.g. temperature, pressure, and flow) that maximize film quality at minimal cost. Typically, this is accomplished by coating a substrate (e.g. a silicon wafer) under a given set of conditions, measuring the film properties ex situ, and adjusting the conditions to improve the film quality. This activity can consume significant time and resources, especially if an additional goal is to achieve uniform films across a large substrate. Process development can be accelerated and economized using in situ measurements. For instance, quartz crystal microbalance (QCM) measurements can be employed to monitor film thickness in real time as the deposition conditions are varied. However, this still requires the careful attention of a skilled operator to make informed choices based on experience and intuition. An alternative strategy is to use machine learning (ML) to analyze the QCM data and adjust the growth conditions based on an algorithm. To explore this possibility, we used ML to optimize the atomic layer deposition (ALD) of Al2O3 with trimethyl aluminum (TMA) and H2O in a viscous-flow tubular reactor using in situ QCM measurements. We initially developed the ML code using simulated QCM data generated by a 1-D model of ALD transport and reaction. This allowed us to tailor the algorithm to ensure saturation of the TMA and H2O ALD reactions and to converge efficiently on the optimal dose and purge times. An additional benefit of these simulations was that we could explore the effects of non-ideal behavior such as a CVD component to the surface reactions and strong interaction between the reaction products and the surface. Next, we interfaced the ML code to our ALD system and allowed the algorithm to optimize the TMA and H2O timings. We observed rapid convergence, as predicted by our simulations, and found that the ML algorithm was capable of adapting to large variations in the initial conditions such as the precursor partial pressures and the carrier gas flow rate. We are now building an array of QCM sensors to measure the thickness simultaneously at 10 locations along our flow tube, and we hope to report on ML opimization of thickness and uniformity using this array.