DEVELOPMENT OF DEEP LEARNING MODELS WITH SYNTHETIC DATA AUGMENTATION FOR THE IDENTIFICATION OF MULTIPLE-TIP ARTIFACTS IN ATOMIC FORCE MICROSCOPY IMAGES

Atomic force microscopy (AFM) is a surface characterization technique capable of acquiring images with nanometric resolution and, under specific conditions, atomic resolution, without compromising sample integrity. Its applications span from materials science to the biological sciences, where AFM is widely employed in the topographic characterization of biomolecules, membranes, cells, and other biophysically relevant nanosystems. As with other imaging techniques, AFM is susceptible to image artifacts, features not representative of the actual sample topography that arise from experimental or instrumental factors. The identification of such artifacts poses a significant barrier to reliable AFM image analysis, requiring extensive technical expertise and interpretive skill developed over years of practice. Even experienced analysts remain subject to perceptual errors and biases, making the process inherently fallible. In this context, machine learning-based automation offers a more consistent and bias-free alternative, capable of standardizing artifact identification and reducing dependence on individual judgment. This work proposes a deep learning model for the automatic identification of multiple-tip artifacts, distortions caused by tip wear or contamination that manifest as duplications or repetitive patterns in AFM images. A dataset of 2,104 images acquired over the past decade by the Laboratório de Física Biológica was labeled into two categories: with artifact and without artifact. Through manual curation, 40 images containing artifacts were identified, composing the real training set. Given the limited size of this set, data augmentation strategies were employed to expand the training data: classical augmentation applied geometric transformations and contrast adjustments to the original images, while synthetic augmentation submitted artifact-free images to a multiple-tip simulation filter to generate artificial positive examples. Prior to training, all images underwent preprocessing for data normalization. Three configurations were trained and compared, without augmentation, with classical augmentation, and with synthetic augmentation, and performance metrics were extracted from the confusion matrix to investigate the comparative efficacy of each strategy and quantify model discriminative capacity. The models trained with synthetic data are expected to generalize to real images, classifying artifact presence with high accuracy and reducing dependence on specialized judgment, representing an advancement toward reliable AFM analysis of biological and biophysical samples.