Prediction of dissolution and nucleation in silicates using
machine learning force fields
Name
Roy-swagata-phd-dmse-2026-thesis Swagata Roy.pdf
Description
Thesis PDF
Size
6.06 MB
Format
Adobe PDF
Checksum (MD5)
1da76cf4ee2d0dbe74bdec3ea8a024be
Author(s)
Roy, Swagata
Advisor(s)
Gómez-Bombarelli, Rafael
Date Issued
February 2026
Publisher
Massachusetts Institute of Technology
Abstract
Understanding how silicates react in water is crucial in a range of fields, from geology and cement chemistry to the synthesis of zeolite catalysts and precipitated silica. Although reactive simulations offer powerful insights at the molecular level, balancing accuracy and scale in these condensed-phase calculations remains a significant challenge. With the advent of machine-learned potential for materials, large-scale calculations seem achievable with high accuracy. In this thesis, machine learning and data-driven analyses are extended to capture complexities inherent to silica gelation, which were then applied to advance the mechanistic understanding of silica polymerization under different environmental conditions using molecular dynamics and kinetic Monte Carlo simulations. We developed a robust reactive machine-learned potential to accurately capture silicate-water reactivity. To ensure the robustness of the model, we introduce a new and general active learning strategy based
on the attribution of the model uncertainty that automatically isolates uncertain regions of bulk simulations to be calculated as small-sized clusters. The potential reproduces the static and dynamic properties of liquid water and solid crystalline silicates, despite having been
trained exclusively on cluster data. We applied the potential to obtain energy barriers for the formation of silica oligomers of any shape and size. Using this, we used a kinetic Monte Carlo study to study the effect of pH and temperature on the dynamics evolution of amorphous silica gel by identifying small building units leading to precipitated silica or crystalline zeolite depending on different conditions. Realizing limitations in the assumptions involved in the Kinetic Monte Carlo study and to incorporate long-range interactions in the study, we trained a coarse grained potential to replicate the silica polymerization dynamics. The potential was trained successfully by distilling a foundation model and was able to accurately capture the time evolution of silicate clusters but had its own limitations in terms of size and time. In general, this thesis provides a detailed description of the preliminary stages of silica oligomerization that lead to an amorphous gel acting as precursors for precipitated silica or zeolites.
MIT Department
Massachusetts Institute of Technology. Department of Materials Science and Engineering
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
