Mapping the Free-Energy Landscape of MFI Zeolite Nanosheet Assembly with the ChIMES Machine-Learned Potential

The performance of ultrathin MFI zeolite nanosheet membranes is dictated by stacking motifs governed by complex interfacial reactivity. Simulating this assembly is hindered by the accuracy/efficiency trade-off between quantum methods and classical potentials. We address this by developing a reactive machine-learned interatomic potential (ML-IAP) using the Chebyshev Interaction Model for Efficient Simulations (ChIMES) framework.

We used the ChIMES active learning workflow to automate the generation of our potential by building the training set while minimizing DFT calculations. A base model trained on bulk zeolite data captures key structural and dynamic properties and was augmented with targeted DFT-MD simulations of water and small siloxanes to describe the reactive chemistry of hydroxylated interfaces.

The resulting potential was employed to compute the free-energy landscape of a double-nanosheet system. Using steered molecular dynamics (SMD), we constructed the potential of mean force (PMF) as a function of inter-sheet separation and alignment to identify favorable stacking motifs. This work links atomic-scale interactions to mesoscale properties, providing a computational framework to guide the rational design of zeolite membranes.