Enhancing small dataset prediction of silver nanoparticle size with deep learning and Latin hypercube sampling framework

Laboratory experiments often face challenges such as inherent complexities, difficulties in data gathering, high costs, and time-consuming procedures. These constraints typically result in a limited amount of experimental data, leading to calculation issues such as overfitting and underfitting. To address these issues, this study applied an integrated framework of deep learning combined with Latin hypercube sampling (LHS) to enhance prediction models based on small datasets. A case study on size prediction in silver nanoparticle synthesis was used to demonstrate the performance of the developed framework. The LHS technique augments the amount of raw data for model development. Consequently, the original raw data and the data generated from LHS were integrated as training data for the development of a deep learning prediction model. This integrated model improved prediction performance, validated by the validation and test dataset R2 values, which are 0.924 and 0.918, respectively. Additionally, the accuracy of unseen data test results was significantly higher when compared to a model trained on a small dataset, with the value rising from 0.442 to 0.893. The proposed framework enables high-accuracy predictions of silver nanoparticle size using small experimental datasets and other conditions within specified boundaries.