Rational Computational Design of Next-Generation Semiconductors

Challenges of environmental pollution, global energy shortage, and overreliance on fossil fuels can be addressed by innovation in solar technology, such as new absorbers for increasing solar cell efficiency and improved photocatalysts for hydrogen production and CO₂ reduction. Novel semiconductors that show bulk stability, promising optoelectronic properties, defect tolerance, and suitable dopability, are desired as substitutes for candidates presently used in these applications, but the atom-composition-structure space of potential materials is practically infinite and not conducive to brute-force experimentation or computation [1]. This necessitates the use of data-driven strategies combining large computational datasets and state-of-the-art machine learning (ML), prior to experimental validation and discovery.
In this work, we developed a rational virtual materials design strategy powered by high-throughput density functional theory (DFT) computations and a variety of descriptor-based and structure-based ML approaches, and applied it to perform multi-objective optimization and discovery in two broad semiconductor material classes: (a) binary, ternary, quaternary, and multi-nary chalcogenides, from CdTe to Cu(In,Ga)Se₂ to Cu₂ZnSnS₄ and complex alloys therein, and (b) A_lB_mX_n type perovskite-inspired materials, including halides and chalcogenides, in a variety of phases and with complex alloying [2-4]. This strategy involved compiling massive DFT datasets of relevant properties using different semi-local and non-local hybrid functionals, performing multi-fidelity active learning [2,5] to rationally generate new high-fidelity data, and training predictive models to accurately obtain any property of interest directly from the semiconductor composition or structure.
One of the primary innovations of this work is the ability to predict the formation energy and electronic levels of point defects, dopants, and defect complexes in crystalline semiconductors, using crystal graph-based neural network (GNN) models, trained upon tens of thousands of structures from bulk and defect calculations [4,6]. These models are deployed for screening of new materials likely to show “defect tolerance” as well as desired stability for particular dopants. The GNN models are combined with optimization algorithms such as Bayesian optimization and simulated annealing to quickly and accurately perform geometry optimization of completely new bulk and defect structures, drastically cutting down the time for full DFT [6]. Using these approaches, we successfully designed dozens of novel halide and chalcogenide compounds with suitable thermodynamic stability (including entropic contributions), electronic band gaps and edges, optical absorption behavior, intrinsic defect tolerance, ability to be doped, and carrier mobilities, concentrations, and cross-sections, with utility for a variety of optoelectronic applications. All the data and models are released to the community via nanoHUB, a nanotechnology repository housed at Purdue, and are currently guiding rational experimental synthesis and characterization efforts.

References
[1] A. Mannodi-Kanakkithodi, Comput. Mater. Sci. 243, 113108 (2024).
[2] J. Yang et al., J. Chem. Phys. 160, 064114 (2024).
[3] M.H. Rahman et al., "High-Throughput Screening of Ternary and Quaternary Chalcogenide Semiconductors for Photovoltaics”, under review.
[4] M.H. Rahman et al., APL Machine Learning. 2, 016122 (2024).
[5] R. Xin et al., J. Phys. Chem. C. 125, 29, 16118–16128 (2021).
[6] G. Cheng et al., Nat. Commun. 13, 1492 (2022).