Machine learning-based structure-property predictions in silica aerogels.

Area of Science:

• Materials Science
• Computational Modeling
• Machine Learning

Background:

• Silica aerogels possess complex structural features effectively modeled by diffusion-limited cluster-cluster aggregation (DLCA).
• Traditional methods require generating DLCA structures and then simulating their fractal properties, a computationally intensive process.

Purpose of the Study:

• To develop an artificial neural network (ANN) for predicting fractal properties of silica aerogels based on DLCA input parameters.
• To invert the ANN for predicting DLCA parameters to reconstruct silica aerogel networks with a target fractal dimension.
• To address the non-uniqueness issue in reconstructing network structures from fractal dimensions.

Main Methods:

• Development of an artificial neural network (ANN) trained on DLCA parameters and corresponding fractal properties.
• Inversion of the trained ANN using a guided gradient descent approach to predict DLCA parameters for a desired fractal dimension.
• Generation and comparison of DLCA model structures from constrained and unconstrained inversions, including pore-size distribution analysis.

Main Results:

• The ANN accurately predicts the fractal dimension of silica aerogels from DLCA parameters with R² = 0.973.
• The inverted ANN successfully predicts DLCA input parameters for a target fractal dimension.
• Constrained inversion of the ANN predicts DLCA model parameters within a 2% error for a desired fractal dimension, resolving non-uniqueness issues.

Conclusions:

• ANNs offer an efficient alternative to direct simulation for predicting and reconstructing silica aerogel fractal properties.
• The developed method enables precise control over the reconstruction of DLCA network structures for targeted material characteristics.
• The study highlights the potential of machine learning in accelerating materials design and discovery for nanoporous materials.