Multimodal LLM-Assisted Diagnosis of Cryptic Failure Modes in High-Throughput Experimentation: A Case Study in Gold Nanorod Synthesis

Self-driving labs typically optimize experiments within a predefined design space. Yet, high-throughput campaigns can fail for reasons that lie outside that space, including reagent state, stock solution quality, and handling history. Here, we present a case study of a five-plate high-throughput gold nanorod synthesis campaign in which a multimodal large language model (LLM) assisted in diagnosing such cryptic failure modes. Low-cost visual characterization provided rapid assessment of seed and synthesis state, while quantitative UV--vis spectroscopy was used when images were insufficient for diagnosis. The campaign included an abnormal seed batch, a visually plausible synthesis with hidden spectroscopic defects, and a plate-wide failure dominated by weak, scattering-like spectra. The resulting interventions focused on seed preparation and reagent handling rather than the screened silver nitrate and ascorbic acid concentrations. To quantify recovery without relying on a single extracted peak, we defined a UV--vis spectral roughness metric based on local spectral interpolation across the synthesis grid. The final synthesis had the lowest roughness, $R_{\mathrm{spec}}=0.636$ compared with $\sim0.772$ for earlier functional runs, and the strongest longitudinal-band peak. The interaction record also captures model errors, including incorrect quantitative predictions and a spectral extraction error during the campaign. These results suggest that multimodal LLMs can complement autonomous optimization in self-driving labs by helping identify when experimental outcomes are unreliable and by escalating from visual to quantitative characterization when needed.