contact: crek[at]iit.demokritos.gr
This thesis proposes a machine learning (ML) architecture designed to enhance and accelerate molecular simulations for nanoporous materials. The architecture integrates three core ML modules: a physics-guided neural network (PGNN), a physics-informed neural network (PINN), and a Gaussian process optimization algorithm. The PGNN operates at the molecular level, utilizing molecular coordinates as input and generating outputs that include the coordinates and the number of adsorbates. Its loss function is specifically designed to minimize the system’s energy, guided by the fundamental physics equations that govern the molecular interactions. Building on this, the PINN scales the problem to the macro level, enabling applications to real-world scenarios. It achieves this by solving differential equations relevant to key processes such as adsorption and diffusion, thereby validating the results against experimental data. Concurrently, the Gaussian process optimization algorithm focuses on inverse design, aiming to optimize the material’s performance at the molecular level by identifying configurations that maximize desired properties. Together, these modules form a comprehensive framework that bridges molecular-scale simulations with macroscopic applications, offering a robust and efficient approach to understanding and designing nanoporous materials.
References
Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations, https://doi.org/10.1016/j.jcp.2018.10.045
Inverse design of ZIFs through artificial intelligence methods, https://doi.org/10.1039/D4CP02488E