Multimodal and Multifidelity Characterization on Combinatorial Thin Films from Uncharted Chemical Spaces

Backed by a unique suite of combinatorial thin-film deposition setups with access to sulfur and phosphorus sources [1], we have explored the Cu-P-S, Sb-P-S, and Ba-P-S phase diagrams by high-throughput experiments. We can now report the first thin-film synthesis of various semiconductors that show significant promise for optoelectronics applications. An obvious question to ask is: What is the best strategy to capture the essential properties and trends of previously unknown combinatorial samples in a time-efficient, yet sufficiently accurate manner?

I will present our group’s current approach to this matter. The essential points are:

- (experiment/computation multimodality). Reliable determination of certain properties from first principles may relax the need for certain experimental measurements that tend to be time-consuming and/or notoriously ambiguous. Example properties are carrier effective masses, direct vs. indirect band gap, orbital-resolved density of states, chemical identification of point defects

- (technique-specific multimodality). Some techniques are naturally amenable to multi-modal characterization. Some examples from our own research are: a UHV analysis chamber with access to five chemical analysis techniques with automatic mapping; a photoluminescence mapping setup that can switch between steady-state, time-resolved, and imaging mode; an electrical-optical measurement stage where electrical properties can be measured as a function of temperature, illumination intensity, and gas atmosphere. Even on the first-principles computational side, some codes are inherently multi-modal. An example is a recent package to calculate carrier mobility limits in semiconductors. To determine the limits arising from different scattering mechanisms, the package has to calculate electronic and phonon band structures, dielectric constant, and carrier effective masses

- (multifidelity) Some properties tend to vary smoothly across a combinatorial sample (i.e., elemental composition, charge states). Other properties tend to make large jumps at certain elemental compositions or process temperatures (i.e., crystal structure, electrical conductivity). Some measurements require high-quality data acquisition but may not require acquisition over a dense grid of points. Other measurements may need just the opposite. I will present our attempts to combine these multifidelity characterization methods into a coherent workflow.

- (relevance) The choice of characterization method should be directed towards obtaining the properties that most directly affect performance in the desired application area. Therefore, application-specific key performance indicators should be defined. In this respect, I will discuss a recently proposed figure of merit for optoelectronic materials for solar energy applications [2,3].

- (artificial intelligence) AI methods can help at various stages of the characterization process, both in the data analysis phase and in the decision-making phase (which property to measure next)

[1] L. A. Mittmann, A. Crovetto, J. Phys. Mater. 2024, 7, 021002.
[2] A. Crovetto, J. Phys. Energy 2024, 6, 025009.
[3] A. Crovetto, 2024, Arxiv preprint, 10.48550/arXiv.2404.14732.