Seminars & Events

MechE-CSE PhD Thesis Defense: Emily Lin

Program MechE-CSE PhD Advisor Evelyn N. Wang, Professor of Mechanical Engineering, MIT Thesis Title Accelerating atomistic simulations via machine learning-based surrogate models for material design and optimization
Date Wednesday, April 15, 2026 Time 3:30–4:30 PM

Location Campus Location: 3-270
Zoom link: https://mit.zoom.us/j/5400768046

Abstract:

Materials innovation is central to the development of high-performance systems for energy and sustainability applications. However, experimental trial-and-error approaches for identifying materials with desired properties are labor-intensive and time-consuming. Computational simulations, which are traditionally used to support material characterization, also exhibit several key limitations, including high computational cost, long runtimes, and sensitivity to hyperparameters. These challenges are further amplified by the vast and complex materials design space. Machine learning (ML) surrogate models offer a promising alternative by bypassing computationally expensive simulations and accelerating material design and optimization. In this thesis, we develop and demonstrate ML surrogate models that enable efficient design across multiple scales, spanning the quantum, atomistic, and application levels.

First, we developed RelaxNet, the first end-to-end neural ODE-based model capable of predicting the coarse-grained relaxation pathway (including final relaxed structure), along with corresponding energy and force field information, given only the initial unrelaxed structure. By starting at the electronic scale, we accelerated structure relaxation processes that are traditionally performed via density functional theory (DFT). We demonstrated the model’s capability to produce physically-consistent, robust, and accurate results at high speed, reducing simulation time from minutes/hours per structure to only minutes per hundreds of structures.

Next, we can use these outputs (e.g., relaxed structure, force fields) to explore molecular simulation in depth, specifically Grand Canonical Monte Carlo (GCMC). Here, we developed two fast, robust, and accurate surrogate models for adsorption property (uptake, heat of adsorption) prediction and full isotherm reconstruction. We developed (1) IsothermNet, a graph-based model for instantaneous property predictions, and (2) IsothermODE, a neural ODE-based model for smooth isotherm reconstruction, interpolation, and extrapolation, with uncertainty quantification. We show that both models decrease full isotherm recovery cost from 20 hours per structure to only seconds per hundreds of structures.

Finally, to comprehensively understand these adsorption outputs in relation to material structure and isotherm behavior/shape, we formulated unifying physical and analytical descriptors that can be used synergistically with our existing ML surrogate models. We show fast isotherm classification and holistic understanding of the adsorbent design space, ultimately allowing for efficient material tuning.

This thesis shows that ML surrogate model and descriptor development can greatly expedite multiscale design, eventually allowing us to translate materials into real-world devices, from catalytic reactors that pretreat polluted flue gas to water harvesters for potable water generation.

Thesis Committee Members:

  • Prof. Evelyn N. Wang, Department of Mechanical Engineering, MIT (Thesis Advisor)
  • Prof. Sili Deng, Department of Mechanical Engineering, MIT (Thesis Co-Advisor)
  • Prof. Rafael Gómez-Bombarelli, Department of Material Science & Engineering, MIT
  • Prof. Nicolas Hadjiconstantinou, Department of Mechanical Engineering, MIT