ICMCTF 2026 Session CM3-1-MoA: Data-Driven Thin Film Design: High-Throughput Experimentation, Simulation, and Machine Learning I

The laws of thermodynamics are often used to predict outcomes of materials synthesis. However, thermodynamics alone cannot explain why some materials are easier and some much harder to grow, or why some systems exhibit strong tendencies toward forming stable amorphous phases and others do not. The situation is much more complicated for non-equilibrium synthesis methods (e.g. vacuum deposition), which often produce states that are not the thermodynamic ground states. In this talk I will describe the computational, first-principles technique we have developed to predict the likelihood for experimental realization of different states. We have found that many of the observed synthesis outcomes across material systems and growth methods can be explained using probabilistic arguments. The critical quantities are the sizes (“widths”) of various local minima on the potential energy surface representing the states available to the system. By measuring the “widths” of local minima using the first-principles random structure sampling we have found that wider the local minimum (more probable in the random sampling) higher are its chances for experimental realization.[1] [#_edn1] This hypothesis has been validated against experiments and used to predict and explain synthesis outcomes more broadly. Examples include our recent work explaining why many layered ternary nitrides preferentially adopt a disordered rocksalt structure in as-grown thin-films despite the existence of ordered ground states much lower in energy.[2] [#_edn2] Or why Y₂WN₄ exhibits a strong tendency toward forming amorphous films, which are shown experimentally to be particularly resistant to crystallization.[3] [#_edn3] I will also discuss our approach to modeling kinetics of structural transformations and its utility in predicting the results of post-growth annealing.

[1] [#_ednref1] V. Stevanovic, Phys. Rev. Lett. 116, 075503 (2016)

[2] [#_ednref2] A. Zakutayev, M. Jankousky, L. Wolf, Y. Feng, C. L. Rom, S. R. Bauers, O. Borkiewicz, D. A. LaVan, R. W. Smaha, and V. Stevanovic, Nat. Synth. 3, 1471 (2024)

[3] [#_ednref3] O. V. Pshyk, S. Zhuk, J. Patidar, A. Wieczorek, A. Sharma, J. Michler, C. Cancellieri, V. Stevanovic, S. Siol, Adv. Mater. 2501074 (2025)

In recent decades, there has been a growing trend in the use of experimental and simulation-based screening methods across various fields of Materials Science. The main goal of these methods is to reduce the time and costs associated with laboratory experimentation. In the area of thin film technologies, especially regarding Physical Vapor Deposition (PVD) methods, combinatorial sputter deposition has emerged as a well-established experimental technique. However, ab initio methods, such as Density Functional Theory (DFT), often face limitations in accurately predicting essential properties like mechanical properties and the solubility of alloy components. The latter is of particular importance when novel thin film materials, such as carbides, nitrides, or borides, are engineered, as the extreme cooling rates during PVD promote the formation of metastable, often uncharted, solid solutions. Here trustful predictions in unexplored phase spaces, would be very helpful to optimize experimental work. However, one reasons why ab initio techniques frequently fail to yield satisfactory results for PVD thin films is that the resulting phase diagrams are markedly different compared to their equilibrium states – typically obtained from CALPHAD.

Nevertheless, over a decade ago, a model was proposed that connected combinatorial sputter deposition experiments with data obtained from DFT and CALPHAD. This model was capable of incorporating the effects of substrate temperature, target power and residual stress in the films. According to literature, the model has proven to be of adequate predictive accuracy, in the case of the metallic W_1-xCu_x and V_1-xCu_x [1] and the ceramic Ti_1-xAl_xN and V_1-xAl_xN [2,3] systems.

The present study aims to extend the methodology by virtue of two principal factors. Firstly, it negates the necessity for existing CALPHAD databases. Secondly, it incorporates the effects of interfacial energies stemming from microstructural causes, such as preferred crystal orientation and grain geometry. The present study focuses on ceramic thin films, and the initial stage involves a concise analysis of model dependencies regarding data from experiment and DFT. Subsequently, a reproduction and extension for the Ti_1-xAl_xN and V_1-xAl_xN systems is presented, with an outlook and presentation of ongoing work to test the method for established carbide and boride systems.

[1] Chang K., et al. Sci. Technol. Adv. Mater. 2016;17:210.

[2] Liu S., et al. Acta Mater. 2019;165:615.

The lecture will cover different ways how to support the experimental research in the field of functional coatings by computer simulations. Various levels of theory (ranging from solid-state physics through atomic-scale ab-initio simulations to atomic-scale simulations based on empirical interaction potentials) and various simulation algorithms (ranging from static calculations of properties through searching for a local energy minimum to reproducing the time evolution of growing films) will be considered. Special attention will be paid to recent developments of the methodology. In all cases, the results will be compared to the experiment.

The first part will deal with a design of multilayered VO₂-based thermochromic coatings for smart energy-saving windows. Integral luminous transmittance and modulation of integral solar energy transmittance will be optimized in parallel, instead of a tradeoff between them. Prediction of coating color will be mentioned as well.

The second part will deal with an identification of maximum N content in amorphous Si–B–C–N networks of various compositions, assuming that it is limited by the formation, presence and loss of N₂ molecules. For example, the difficulties with the preparation of C₃N₄ will be explained by a maximum achievable stable N content in CN_x of 42%.

The third part will deal with reproducing the atom-by-atom growth of functional coatings in a wide range of conditions, using examples such as ZrO₂, Cu-Zr or Ti-Al-N. Effects of energy distribution function (not only average energy delivered into the films), mass of arriving particles (momentum delivered into the films) and growth temperature will be explained.

Advanced device integration requires process-aware material descriptors that capture how precursor chemistry, plasma species and growth kinetics determine thin-film properties. We present a multiscale pipeline that couples DFT energetics, an active-learning M3GNet interatomic potential, automated PES exploration, and 3D kinetic Monte-Carlo growth simulations to produce spatially resolved property descriptors (growth rate, composition maps, defect/trap proxies, band-gap and dielectric indicators).

Starting from a foundation M3GNet universal machine learning interatomic potential (MLIP) and a small DFT seed set, we fine-tune and use uncertainty-guided sampling to automatically find intermediates and transition states; high-uncertainty configurations are re-computed with DFT and fed back to the MLIP, accelerating accurate exploration of the Potential Energy Surface (PES) at a little over MLIP cost with near-DFT accuracy. The verified reaction network parameterizes kMC to predict thin flim formation and etching, as functions of precursor, pulse timing, temperature and flux. We demonstrate the workflow on a Ru–H ALE case study: ML-accelerated PES exploration uncovered dissociative channels that shift band-gap and fixed-charge proxies; kMC maps reveal process windows minimizing interface trap formation while preserving selectivity. The approach provides compact, validated descriptors for process tuning and device correlation, shortening R&D cycles and guiding targeted experiments.