Quantitative analysis of the prediction performance of a Convolutional Neural Network evaluating the surface elastic energy of a strained film

TL;DR

This paper develops a deep learning model to accurately predict surface elastic energy density in strained films, demonstrating high accuracy, robustness, and generalization to new geometries and domain sizes.

Contribution

It introduces a fully convolutional neural network trained on a large dataset of surface profiles and elastic energy densities, achieving precise predictions and assessing generalization beyond the training data.

Findings

Neural network accurately predicts elastic energy density profiles.

Model generalizes well to unseen surface features and geometries.

Effective in time-integration of surface evolution PDEs.

Abstract

A Deep Learning approach is devised to estimate the elastic energy density $\rho$ at the free surface of an undulated stressed film. About 190000 arbitrary surface profiles h(x) are randomly generated by Perlin noise and paired with the corresponding elastic energy density profiles $\rho(x)$, computed by a semi-analytical Green's function approximation, suitable for small-slope morphologies. The resulting dataset and smaller subsets of it are used for the training of a Fully Convolutional Neural Network. The trained models are shown to return quantitative predictions of $\rho$, not only in terms of convergence of the loss function during training, but also in validation and testing, with better results in the case of the larger dataset. Extensive tests are performed to assess the generalization capability of the Neural Network model when applied to profiles with localized features or assigned geometries not included in the original dataset. Moreover, its possible exploitation on domain sizes beyond the one used in the training is also analyzed in-depth. The conditions providing a one-to-one reproduction of the ground-truth $\rho(x)$ profiles computed by the Green's approximation are highlighted along with critical cases. The accuracy and robustness of the deep-learned $\rho(x)$ are further demonstrated in the time-integration of surface evolution problems described by simple partial differential equations of evaporation/condensation and surface diffusion.