Quantum 2025 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
MON: Monday Contributed Sessions
MON 5: Optical Quantum Devices
MON 5.7: Talk
Monday, September 8, 2025, 15:45–16:00, ZHG006
Leveraging Large Language Models as Qualitative Figures of Merit — •Liam Shelling Neto1, Nico Wagner1, and Stefanie Kroker1,2 — 1Technische Universität Braunschweig, Institute of Semiconductor Technology, Hans-Sommer-Str. 66, Braunschweig, 38106, Germany — 2Physikalisch-Technische Bundesanstalt, Bundesallee 100, Braunschweig, 38116, Germany
Inverse design of photonic components, particularly in quantum technology, typically relies on quantitative figures of merit (FoMs) such as efficiency, transmission, or beam waist, metrics derived from precisely defined simulation data. However, many design goals are inherently qualitative or difficult to express numerically. We propose using large language models (LLMs) as evaluators of qualitative figures of merit (qFoMs) to complement traditional FoMs. These qFoMs can assess visual or descriptive attributes of optical responses, such as field distributions, via natural language prompts and image inputs. For example, when optimizing a grating coupler to produce a Gaussian beam, early designs often yield patterns far from Gaussian, rendering quantitative fits ineffective. An LLM-based qFoM, similar to a human, can still assign a "Gaussianity" score to guide the optimizer in the right direction, even in these early, low-performance stages. By integrating qFoMs with traditional metrics, LLMs act as pseudo-intelligent agents that bridge the gap between human intuition and algorithmic evaluation, enabling more robust and flexible optimization, especially when the target functionality is poorly defined or initially unmet.
Keywords: Inverse Design; Photonics; Large Language Models; Qualitative Optimization; Hybrid Figures of Merit