Nathan Stern, an analytical chemist at Amway / Nutrilite (Ada, USA), specializes in development and validation of various types of chromatographic test methods. He develops both quantitative (UHPLC) and qualitative (HPTLC) phytochemical tests that are designed to ensure ingredient quality and authenticity. He is also experienced in using machine learning techniques, especially as they apply to analytical chemistry.

The primary goal of Nathan’s research was to develop a machine learning system to automate the evaluation of HPTLC-generated botanical fingerprints for the determination of botanical identity. The current industry approach involves manually comparing these HPTLC images with botanical reference materials of both authentic species and common adulterants. Machine learning techniques, such as machine vision, can enable faster and more accurate identification of botanicals.

All HPTLC images were obtained from the online and publicly available HPTLC Association Atlas repository. This includes 77 total image files for the following species: Alpinia officinarum, Boesenbergia rotunda, Kaempferia galanga, Kaempferia parviflora, Zingiber montanum, Zingiber officinale, and Zingiber zerumbet [2].

To 1.0 g of each powdered sample 10 mL of methanol are added, followed by 10 minutes of sonication. The samples are centrifuged, and the supernatant is used as test solution [2].

HPTLC plates silica gel 60 F254 (Merck), 20 × 10 cm are used.

2.0 μL of sample and standard solutions are applied as bands with the Automatic TLC Sampler (ATS 4), 15 tracks, band length 8.0 mm, distance from the left edge 20.0 mm, track distance 11.4 mm, distance from lower edge 8.0 mm [2].

Plates are developed in the Automatic Developing Chamber (ADC 2) with chamber saturation (with filter paper) for 20 min and after activation for 10 minutes at a relative humidity of 33% using a saturated magnesium chloride solution, development with toluene – ethyl acetate 3:1 (V/V) to the migration distance of 70 mm (from lower edge), followed by drying for 5 min [2].

The plates are derivatized using an anisaldehyde reagent prepared by adding 10 mL of acetic acid, 5 mL of sulfuric acid, and 0.5 mL of anisaldehyde to 85 mL of icecooled methanol. The derivatization reagent (3 mL) is sprayed using the Derivatizer (blue nozzle, level 3). The plates are heated at 100°C for 3 min and allowed to cool before detection [2].

Images of the plates are captured with the TLC Visualizer 3 in UV 254 nm, UV 366 nm, and white light after development, and again after derivatization in UV 366 nm and white light.

The machine vision model was created on a computer system utilizing an Nvidia GeForce RTX 3070 GPU for computation. The machine learning software system was implemented in Python using Visual Studio Code as the IDE and PyTorch as the machine learning framework. The machine vision system is comprised of several different neural networks, including a deep conditional generative adversarial network (DCGAN) made up of a discriminator and a generator, as well as a deep convolutional neural network (deep CNN).

The role of the DCGAN in the system is to augment the limited dataset by creating a large number of synthetic HPTLC images for each species, based on a partition of real HPTLC image data. This synthetic data was then used to train the deep CNN model, which was validated separately against both real and synthetic HPTLC image datasets.