Improving the prediction of bitumen’s density and thermal expansion by optimizing artificial neural networks with Optuna and TensorFlow

Abstract

Previous work demonstrated that Random Forest Regressors (RFRs) could estimate the physical properties of bitumen using molecular descriptors derived from Molecular Dynamics (MD) simulations, thereby reducing the need for computationally intensive simulations. However, due to their decision-tree structure, RFRs lack true predictive capabilities, particularly for interpolation and extrapolation beyond the training data.

This study advances that foundation by employing Artificial Neural Networks (ANNs), which—when properly trained—can capture complex relationships with greater continuity and generalizability. Beyond simply replacing RFRs, we develop a fully automated framework for constructing Machine Learning Models (MLMs) to predict density and thermal expansion coefficients of bitumen. Using Optuna for hyperparameter optimization, we ensure that the information extracted from MD simulations is utilized effectively.

The resulting ANN models accurately reproduce MD-predicted densities, achieving R²>0.99, MSEs below 0.1 %, and maximum absolute errors below 5 % on test data. In addition to reducing computational cost, the models exhibit improved interpolation and extrapolation capabilities, enabling reliable predictions for properties, ranges, and compositions not explicitly simulated.

Key aspects of our approach include:

• •
  Transitioning from RFRs to ANNs, improving generalization, interpolation, and predictive accuracy.
• •
  Automated hyperparameter optimization, leveraging Optuna to maximize model efficiency.
• •
  Expanding applicability, enabling property prediction for unseen compositions without additional MD simulations.