Digital Fingerprinting of Complex Liquids Using a Reconfigurable Multi-Sensor System with Foundation Models

Area of Science:

• Analytical Chemistry
• Artificial Intelligence
• Machine Learning

Background:

• Chemical sensor arrays combined with machine learning offer advanced sensing capabilities beyond traditional methods.
• Portable sensor systems generate limited data, hindering the training of large machine learning models for chemical sensing.
• Foundation models excel at zero-shot learning across various data types, showing potential for transfer learning.

Purpose of the Study:

• To develop a generalizable framework for AI-assisted chemical sensing using foundation models.
• To create effective data representations for chemical sensor signals.
• To enable accurate chemical sensing with limited domain-specific training data.

Main Methods:

• A novel framework was developed to translate signals from simple, portable multi-sensor systems into visual fingerprints.
• Pretrained vision models were incorporated into a pipeline for chemical sensing tasks.
• The approach was tested on four unrelated chemical sensing tasks with limited training data.

Main Results:

• The pipeline successfully translated sensor signals into visual fingerprints of liquid samples.
• An average classification accuracy was achieved across four diverse chemical sensing tasks.
• The performance matched or surpassed expert-curated sensor signal features.

Conclusions:

• Transfer learning from foundation models provides a generalizable approach for AI-assisted chemical sensing.
• This method enhances data processing for ease-of-use and broad applicability in generic sensing applications.
• The approach overcomes limitations of data-scarce portable sensor systems.