Analyzing Visualization Practices in Modeling and Simulation

As data becomes more complex, it is imperative that it is communicated in a manner that is equally advanced. Poorly made figures in papers focused on data collection and representation can obscure meaning and mislead readers, pointing to an opportunity for the field to evolve its visual standards alongside its models. Prior to advancements in computer graphics and specialty tools, data visualization was limited by technological constraints. This has changed with the emergence of more powerful, affordable computers, giving researchers more diverse, polished figures to support their findings. Digestible visualizations are increasingly important as researchers aim to reach wider audiences, especially in modeling and simulation. This study evaluates visualization quality over the span of 60 years to understand the evolution of data visualization practices and investigate potential challenges. We cover seven prominent journals and conferences in the field of modeling and simulation, analyzing over ten thousand papers published between 1963 and 2025. First, we utilized and fine-tuned a transformer-based document layout Artificial Intelligence (AI) model to locate and extract figures and metadata from these papers. We then trained a couple of AI classifiers to distinguish images of data visualizations from other images using a training-test-validation set that we curated. We achieved an accuracy of 99% on our test set using a vision transformer (ViT), improving on the accuracy of 95% that we measured using a convolutional neural network (CNN). The next step is to evaluate figures that are classified as data visualization with respect to their conformity with modern visualization practices. We believe that our results will shed light on the development and current state of the modeling and simulation field with respect to data visualization and offer practical insight to researchers.