Automated Solubility Analysis System and Method Using Computer Vision and Machine Learning

Gahee Kim, Minwoo Jeon, Hyun Do Choi, Jun Ki Cho, Youn-Suk Choi, Hyoseok HwangarXiv 2023 · DOI · arXiv

Introduction

Solubility assessment is a fundamental requirement in chemical synthesis and purification. Traditionally, researchers have relied on manual visual inspection or invasive analytical techniques like High-Performance Liquid Chromatography (HPLC) and spectrometry. However, these methods are often slow, require expensive equipment, and can disrupt the reaction environment. This paper introduces the Automated Solubility Analysis System (ASAS), a non-invasive, real-time monitoring solution that leverages consumer-grade hardware and machine learning to automate the detection of dissolution states.

System Architecture and Hardware

The ASAS utilizes a cost-effective hardware configuration designed for accessibility and integration into automated labs. The setup consists of a smartphone camera (Samsung Galaxy S21) for image acquisition and a tablet display (Samsung Galaxy Tab S7 FE) to provide controlled backgrounds. These components are secured using 3D-printed jigs to maintain a consistent optical path through the sample flask.

The Automated Solubility Analysis System (ASAS) hardware setup and communication protocol between the camera, display, and server.

As shown in the hardware diagram, the system employs a "dual-background" technique. The tablet displays two specific patterns behind the flask: a solid white background and a checkerboard pattern. This approach allows the system to capture different optical phenomena; the white background is ideal for spotting discrete particles, while the checkerboard pattern is highly sensitive to turbidity (cloudiness) caused by light scattering and refractive index changes.

Methodology: Computer Vision and Machine Learning

The core of the ASAS is a multi-stage computer vision pipeline that processes the captured images to extract meaningful physical features. The process begins with the Hough circle transform, which automatically identifies the flask as the Region of Interest (ROI). Once the ROI is defined, the system employs three specialized analysis modules:

  1. Radial Profile Analysis (RPA) & Particle Amount Analysis (PAA): These modules analyze the white background images to detect individual undissolved particles based on intensity gradients and pixel counts.
  2. Superposition Analysis (SA): This module processes the checkerboard images. By measuring the degree to which the grid lines are obscured or distorted—quantified as the "superposition ratio"—the system can accurately determine the turbidity of the solution.
The complete ASAS workflow, including image capture, preprocessing, feature extraction (RPA, PAA, SA), and SVM classification.

Following feature extraction, the data is fed into a Linear Support Vector Machine (SVM) classifier. The model categorizes the solution into one of three states: Pass (clear/dissolved), Fail 1 (turbid/cloudy), or Fail 2 (visible particles).

Experimental Validation and Results

The researchers validated the system using a diverse dataset encompassing both inorganic salts (e.g., CuSO4, PdOAc) in water and organic substances (e.g., 2-bromo-4-phenylpyridine) in various organic solvents. To enhance the robustness of the SVM, the initial training set was expanded via image augmentation to 1,224 images.

In independent testing across 202 cases, the ASAS achieved a 95% classification accuracy. The system proved particularly reliable at identifying "Fail" states, which is critical for preventing equipment malfunctions like chromatograph clogging in automated systems. Furthermore, real-time monitoring experiments demonstrated that the system could successfully track the dissolution process over several hours, accurately identifying the precise moment a solution transitioned from turbid to fully dissolved.

Conclusion and Future Directions

The ASAS provides a scalable, low-cost solution for real-time solubility sensing. By monitoring the entire flask rather than a microscopic area, it offers a more comprehensive assessment of the sample state. While currently a qualitative classifier, future work aims to refine the system's detection limits and integrate it into fully autonomous "closed-loop" synthetic systems. Such an integration would allow the system to automatically trigger solvent addition or temperature adjustments based on real-time solubility data.

Summary updated

BibTeX
@misc{kim2023automated,
  author = {Gahee Kim and Minwoo Jeon and Hyun Do Choi and Jun Ki Cho and Youn-Suk Choi and Hyoseok Hwang},
  title = {{Automated Solubility Analysis System and Method Using Computer Vision and Machine Learning}},
  year = {2023},
  eprint = {2304.12972},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2304.12972},
}
Figure