PhD Thesis Proposal: Gideon Kassa

Optional ZOOM LINK

"Novel Photovoltaics: From simulation to synthesis and characterization"

Abstract

High-throughput (HT) screening and materials generation have recently received attention. This work presents the successful implementation of HT screening using density functional theory (DFT) to identify a promising photovoltaic, BaCd2P2 (BCP), and the development of a novel machine learning (ML) framework designed to expedite HT materials discovery. BCP was identified as a strong candidate for single-junction solar cells through HT screening. It was synthesized using solid-state reactions, and its properties were experimentally investigated. It was found to have a room-temperature direct band gap of 1.46 eV. The impressive performance of BCP is demonstrated by comparing its optoelectronic properties with that of GaAs. Despite its lower purity, BCP exhibits similar band-edge photoluminescence (PL) response and effective open-circuit voltage as a high-quality single-crystal GaAs wafer. Interestingly, BCP shows a notably higher photocurrent than the GaAs wafer. Moreover, BCP has better impurity tolerance than GaAs; this is demonstrated by showing GaAs with similar purity as BCP does not display room-temperature band-edge PL emission. Intrinsic defect calculations using HSE06 further indicate that there is a much lower concentration of the dominant deep nonradiative recombination center in BCP, P_Cd, compared to the analogous defect in GaAs, As_Ga.

These findings highlight the exceptional impurity and intrinsic defect tolerance of BCP and underscore its potential as a cost-effective, high-performance solar absorber. BCP’s excellent performance demonstrates the power of HT screening in identifying new photovoltaic materials. While DFT is currently the preferred method, the computational cost of high-accuracy hybrid functionals restricts the range of materials that can be explored. Consequently, lower-accuracy DFT functionals like PBE (Perdew-Burke-Ernzerhof) are often used, though they remain computationally expensive. To address this, a physics-informed ML model is under development to predict material properties at a fraction of the computational cost. Trained on high-accuracy HSE06 data, this model surpasses the speed and accuracy of DFT functionals like PBE. The framework is a message-passing neural network that models crystal structures as graphs. Additionally, a material property-conditioned Denoising Diffusion Probabilistic Model that uses graph convolutions is being developed to generate novel materials. This model will accelerate the materials discovery process and development of next-generation photovoltaics.

Thesis Committee

• Jifeng Liu (Chair)
• Geoffroy Hautier
• Soroush Vosoughi

Contact

For more information, contact Thayer Registrar at thayer.registrar@dartmouth.edu.