LLM-Assembled Multiscale Cascades for High-Throughput Screening: The Case of Thermoelectric Materials
1Applied Artificial Intelligence Initiative, Deakin University, Geelong, Victoria 3216, Australia.
None:
High-throughput screening workflows often rank materials with a sequence of filters, but a single sequence can hide how strongly the final ranking depends on the chosen physical approximations. Here a multiscale cascade means an ordered workflow in which outputs from electronic, lattice, microstructural, uncertainty, and device-level models are passed from one layer to the next; the layers are theory or surrogate-model steps, not layers of LLMs. We use a large language model (LLM) as a workflow-design assistant to assemble candidate model stacks from the literature, after which the equations, code, and physical handoffs are inspected and implemented by the author. The test case is thermoelectric screening, where the dimensionless figure of merit ZT = S2σT/(κe + κL) combines the Seebeck coefficient S, electrical conductivity σ, electronic thermal conductivity κe, and lattice thermal conductivity κL. Two independently assembled eight-layer cascades are applied to the same 314-compound vacancy-containing chalcogenide library. LLM1 uses Fan-Migdal band gap renormalization, a Kubo-DMFT transport surrogate, and a literature-trained Gaussian process for κL. LLM2 replaces those three early layers with Bose-Einstein band gap renormalization, acoustic-phonon Boltzmann transport, and a Debye-Callaway integral. The common four-stage screen sends 15 compounds to full evaluation in each cascade: LLM1 selects tellurides headed by CuAlTe2 (ZTpeak = 8.61), whereas LLM2 selects nonoverlapping sulfides headed by CuPb2S4 (ZTpeak = 0.325). The absolute LLM1 values are not claimed as validated performance forecasts: CuAlTe2 is literature-supported as a promising bulk thermoelectric, but reported and expected values are closer to ZT ≲ 2 than to 8.6. This study underscores the importance of transparently comparing multiple workflow designs to understand the sensitivity and reliability of high-throughput screening outcomes in materials discovery.
