Multimodal LLM-Assisted Diagnosis of Cryptic Failure Modes in High-Throughput Experimentation: A Case Study in Gold Nanorod Synthesis
- 1Chemical and Biomolecular Engineering
- 2Institute of Technology
- 3Electrical and Computer Engineering
Abstract
Self-driving labs typically optimize experiments within a predefined design space. Yet high-throughput campaigns can fail for reasons outside this design space, in-cluding 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) helped diagnose such cryptic failure modes. We used low-cost camera images for rapid assessment of seed and synthesis state, with UV–vis spectroscopy when the images were insufficient for failure diagnosis. The campaign included an abnormal seed batch, a visually plausible synthesis with hidden spectroscopic defects, and a plate-wide failed synthesis in which most wells showed weak spectra with a strong rise toward the near-infrared. In response, the LLM intervened by questioning the seed preparation step and reagent handling instead of optimizing the screened concentrations of silver nitrate and ascorbic acid. To quantify recovery, we defined a UV–vis spectral roughness metric based on local spectral interpolation across the synthesis grid. The final synthesis had the lowest roughness, Rspec = 0.636 compared with ∼ 0.772 for earlier successful runs, and the highest median longitudinal-window absorbance. Our results show the potential of multimodal LLMs upstream of autonomous optimization in self-driving labs to identify when an experiment has failed for reasons outside the nominal design space and when additional characterization is needed before optimization should continue.
Related
- Yield Prediction and Condition Recommendation on High-Throughput Experimentation Datasets: Encoding Effects and Noise Ceilings
- High-Throughput Experimentation Enables the C6-Functionalisation of 1H-Indazol-3-Amines as Antitubercular Agents
- Combinatorial Steiner Methodology for Efficiently Exploring Chemical Space in CO ₂ Hydrogenation to Fischer-Tropsch Reactivity via High-throughput Experimentation
- Autonomous Experimentation Enables Rapid Optimization of Photochemical Continuous-Flow Synthesis of Diazaxanthilidene Fluorophores
- Dynamic Flow Experimentation for Data-Rich Polymer Synthesis
BibTeX
@misc{patel2026multimodal,
title = {{Multimodal LLM-Assisted Diagnosis of Cryptic Failure Modes in High-Throughput Experimentation: A Case Study in Gold Nanorod Synthesis}},
author = {Riya Patel and Gabriela Briceno and Amirali Aghazadeh and Vida Jamali},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15009205/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009205/v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - Multimodal LLM-Assisted Diagnosis of Cryptic Failure Modes in High-Throughput Experimentation: A Case Study in Gold Nanorod Synthesis
AU - Riya Patel
AU - Gabriela Briceno
AU - Amirali Aghazadeh
AU - Vida Jamali
PY - 2026
DA - 2026/09/21
PB - ChemRxiv
DO - 10.26434/chemrxiv.15009205/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009205/v1
AB - Self-driving labs typically optimize experiments within a predefined design space. Yet high-throughput campaigns can fail for reasons outside this design space, in-cluding 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) helped diagnose such cryptic failure modes. We used low-cost camera images for rapid assessment of seed and synthesis state, with UV–vis spectroscopy when the images were insufficient for failure diagnosis. The campaign included an abnormal seed batch, a visually plausible synthesis with hidden spectroscopic defects, and a plate-wide failed synthesis in which most wells showed weak spectra with a strong rise toward the near-infrared. In response, the LLM intervened by questioning the seed preparation step and reagent handling instead of optimizing the screened concentrations of silver nitrate and ascorbic acid. To quantify recovery, we defined a UV–vis spectral roughness metric based on local spectral interpolation across the synthesis grid. The final synthesis had the lowest roughness, Rspec = 0.636 compared with ∼ 0.772 for earlier successful runs, and the highest median longitudinal-window absorbance. Our results show the potential of multimodal LLMs upstream of autonomous optimization in self-driving labs to identify when an experiment has failed for reasons outside the nominal design space and when additional characterization is needed before optimization should continue.
ER -