Unveiling the Potential of AI for Nanomaterial Morphology Prediction

TL;DR

The paper addresses the challenge of predicting nanomaterial morphology under realistic data constraints by building a large multimodal calcium carbonate dataset and evaluating classical ML, large language models, and a text-to-image pipeline. It demonstrates that tree-based ensembles achieve strong shape accuracy (≈0.80) and that GPT-4 and related LLMs can match or exceed these results in few-shot settings, offering a language-driven route to morphology prediction. A three-component text-to-image prototype further explores mapping synthesis descriptions to nanoparticle images, showing potential for in silico morphology visualization albeit limited by data; the work also exposes substantial needs for larger, standardized datasets and improved SEM-data handling. Overall, the study provides a rigorous benchmark and showcases promising AI approaches for nanomaterial design, potentially reducing experimental overhead and accelerating discovery. The findings emphasize that combining statistical feature analyses, classical ML, and LLM-driven prompts can yield robust morphology predictions with practical implications for materials science workflows.

Abstract

Creation of nanomaterials with specific morphology remains a complex experimental process, even though there is a growing demand for these materials in various industry sectors. This study explores the potential of AI to predict the morphology of nanoparticles within the data availability constraints. For that, we first generated a new multi-modal dataset that is double the size of analogous studies. Then, we systematically evaluated performance of classical machine learning and large language models in prediction of nanomaterial shapes and sizes. Finally, we prototyped a text-to-image system, discussed the obtained empirical results, as well as the limitations and promises of existing approaches.

Figures (6)

• Figure 1: Average accuracy of GPT-4 for different number of samples in prompt taken from the training set. Sampling method: only target classes in prompt. Syntheses presented in the textual format. Colors correspond to different shapes of nanoparticles.
• Figure 2: A schematic of the text-to-image system prototype. A) VAE training. The images of nanoparticles are used to train a variational autoencoder (VAE). B) Final model inference. The corresponding synthesis procedures are converted into vector representations with a pretrained BERT (bottom left). The “linking” autoencoder is trained to map text and image representations (bottom center). Finally, the decoder of the VAE is used to generate new images of nanomaterials based on the descriptions of syntheses (bottom right).
• Figure 3: A) PSNR and SSIM metrics of the selected VAE by epoch. B) Training loss of the selected VAE by epoch.
• Figure 4: Results of feature importance analysis in the form of SHAP values for the top 10 features for the best Random Forest model for predicting stick shaped nanoparticles.
• Figure 5: A) The encoder-decoder architecture of the linking neural network. B) Comparison of real images to their VAE reconstructions and the images generated from the corresponding synthesis texts.