Firefly-Geni: Automated Iteration-Guided Generative Discovery of Thermally Activated Delayed Fluorescence Molecules
Year: 2026
Type: article
Abstract: The efficient discovery of thermally activated delayed fluorescence (TADF) molecules remains heavily dependent on expert-driven design rules, while the vast structure–property knowledge accumulated in the literature stays largely fragmented and underutilized. Here, we present Firefly-Geni, an automated and interpretable active-evolution framework that transforms dispersed literature knowledge into structurally novel TADF candidates through a closed-loop workflow. Firefly-Geni integrates large-language-model (LLM)-assisted data extraction, multitask property prediction, conditional molecular generation, and theoretical evaluation. Its data-mining module converts unstructured literature into a curated database of experimental property entries. A dual-molecular-graph multitask predictor and a conditional generator are integrated with quantum-chemical feedback, enabling iterative refinement toward structural novelty. From 100,000 molecules generated under target conditions of a singlet–triplet energy gap ΔEST = 0.05 eV and a synthetic accessibility (SA) score of 2.5, Firefly-Geni identifies 100 structurally novel candidates computationally predicted to support reverse intersystem crossing and to be synthetically tractable, spanning diverse emission wavelengths. One synthesized candidate exhibited an experimental ΔEST of 0.019 eV and a film photoluminescence quantum yield (PLQY) of 84% under nitrogen, broadly consistent with the predicted small gap and high PLQY. Interpretability analyses further reveal chemically meaningful structure–property relationships and generation preferences, confirming that the candidates encode physically relevant low-ΔEST design motifs. This work offers a closed-loop, explainable strategy for actively evolving TADF candidates from literature knowledge, providing a generalizable route for data-driven molecular discovery.
Source: ACS Central Science
Institutions: Donghua University, University of Wollongong
Language: English
Subfield: Materials Chemistry
Field: Materials Science
Domain: Physical Sciences
Keywords: thermally activated delayed fluorescence, active learning (machine learning), generative molecular design, large language models, literature data extraction, structure-property relationships, reverse intersystem crossing, singlet-triplet energy gap, photoluminescence quantum yield, synthetic accessibility
Open Access status: diamond
Awards: G408509690, G7560259917
OpenAlex ID: W7167016259
