PervoTransformer: Multi-Scale Transformer-Based Device Design and Atomistic Modeling of Perovskites
Year: 2026
Type: preprint
Abstract: Perovskite solar cell design involves a large and highly coupled space of material compositions, layer structures, and physical parameters, making systematic exploration of complete device architectures computationally challenging. This work introduces PervoTransformer, a transformer-based framework that represents complete device architectures as sequences and learns device-level behavior directly from this representation. A RoBERTa model is fine-tuned on SCAPS-generated device sequences spanning six perovskite families, including MAPbI3, FAPbI3, CsPbI3, and their Sn-based analogs. The model simultaneously predicts multiple performance metrics, including power conversion efficiency (PCE), short-circuit current density (Jsc), open-circuit voltage (Voc), and fill factor (FF), demonstrating the ability of transformer encoders to capture coupled, nonlinear dependencies across complete device stacks. To explore configurations beyond the training distribution, a GPT-2 model is fine-tuned using the same device-sequence representation to generate candidate architectures with new combinations of compositional and physical parameters. Generated configurations are screened for validity and uniqueness and subsequently evaluated using the fine-tuned RoBERTa model, enabling efficient exploration of under-sampled regions of the device design space. In parallel, machine-learned interatomic potentials based on the MACE framework are fine-tuned for CsSnX3 (X = Cl,Br, I) using available DFT data, enabling efficient structural relaxation and exploration of lead-free Sn-based perovskites. The resulting multiscale framework combines sequence-based transformer models for device-level design with machine-learned interatomic potentials for atomistic materials exploration, while retaining first-principles calculations for detailed physical validation. This approach provides a scalable strategy for navigating the coupled materials and device design spaces of perovskite solar cells while reducing reliance on computationally expensive simulations during initial exploration.
Source: ChemRxiv
Language: English
Subfield: Materials Chemistry
Field: Materials Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7214789780