Material-AI
Material-AI
Embedded Files
Deep Learning is started to make an impact in physical science. Elaborating domain knowledge in material simulation and design, I have been working on the research projects in the intersection of AI and material science. The projects I have been working on so far includes AI driven drug discovery, AI driven small molecular design, developing neural-net based surrogate functions to emulate physical simulation, such as DFT (density functional theory), and MD (molecular dynamics). Specifically, I have led DOE-funded project to optimize molecular structure using reinforcement learning by including physical simulation (DFT in his case) in a loop. Several works has been presented in scientific journals and ML conferences.
Deep Learning is started to make an impact in physical science. Elaborating domain knowledge in material simulation and design, I have been working on the research projects in the intersection of AI and material science. The projects I have been working on so far includes AI driven drug discovery, AI driven small molecular design, developing neural-net based surrogate functions to emulate physical simulation, such as DFT (density functional theory), and MD (molecular dynamics). Specifically, I have led DOE-funded project to optimize molecular structure using reinforcement learning by including physical simulation (DFT in his case) in a loop. Several works has been presented in scientific journals and ML conferences.
Optimizing 3D structure of small molecules using reinforcement learning.
1. Conventional methods to predict 3D molecular structure are based on iterative stochastic optimization techniques by calculating energy using physics-based electronic structure modeling such as Density Functional Theory (DFT) or Molecular Dynamics (MD) which is computationally expensive and multi-modal by in nature.
2. As a cost-efficient and relatively deterministic alternative, we propose a novel RL-based algorithm to optimize 3D structure of single molecule based on DDPG method.
3. Our model teaches the agent to find the best movement trajectory of each atom to reach the correct structure by guiding their movement using three reward approaches based on DFT calculation.
4. Our experiment shows that our model successfully predicts the most energetically stable 3D structure of small aromatic-hydrocarbon molecules.
Optimizing 3D structure of small molecules using reinforcement learning.
1. Conventional methods to predict 3D molecular structure are based on iterative stochastic optimization techniques by calculating energy using physics-based electronic structure modeling such as Density Functional Theory (DFT) or Molecular Dynamics (MD) which is computationally expensive and multi-modal by in nature.
2. As a cost-efficient and relatively deterministic alternative, we propose a novel RL-based algorithm to optimize 3D structure of single molecule based on DDPG method.
3. Our model teaches the agent to find the best movement trajectory of each atom to reach the correct structure by guiding their movement using three reward approaches based on DFT calculation.
4. Our experiment shows that our model successfully predicts the most energetically stable 3D structure of small aromatic-hydrocarbon molecules.
Selected Publications
Bhavya Kailkhura, Brian Gallagher, Sookyung Kim, Anna Hiszpanski, Yong T. Han. "Reliable and explainable machine-learning methods for accelerated material discovery". Nature, NPJ. Computational Material, 2019.
Workshops (* Equal contribution)
Sookyung Kim*, Yongwoo Cho*, Peggy Li, Mike Surh, Yong T. Han. "Physics-guided Reinforcement Learning for 3D Molecular Structures". ML for Physical Science on NeurIPS (Spotlight Talk), 2019.
Sookyung Kim, Peggy Li, Joanne Kim, Piyush Karende, Yong T. Han. "Optimizing 3D structure of H2O molecule using DDPQ". Reinforcement Learning for Real Life Workshop on International Conference on Machine Learning (ICML), 2019.
Developing Physics-informed DFT(Density Functional Theory) surrogate function
1. Estimating the energetic properties of molecular systems is a critical task in material design. With the trade-off between accuracy and computational cost, various methods have been used to predict the energy of materials, including recent neural-net-based models.
2. However, most existing neural-net models are context-free (physics-ignoring) black-box models, limiting their applications to predict energy only within the distribution of the training set and thus preventing from being applied to the real practice of molecular design.
3. Inspired by the physical mechanism of the interatomic potential, we propose a physics-driven energy prediction model using a Transformer.
4. Our model is trained not only on the energy regression in the training set, but also with conditions inspired by physical insights and self-supervision based on Masked Atomic Modeling, making it adaptable to the optimization of molecular structure beyond the range observed during training, taking a step towards realizable molecular structure optimization.
Selected Publications
Seunghoon Yi, Hongkee Yoon, Youngwoo Cho, Seungwoo Kho, Sookyung Kim and Joonseok Lee. "Physics-empowered Molecular Representation Learning". Submitted to ICLR-2023
Workshops (* Equal contribution)
Youngwoo Cho, Hongkee Yoon, Seunghun Yi, Joonseok Lee, MJ Han, Jaegul Choo, Sookyung Kim, "Deep-DFT: A Physics-ML Hybrid Approach to Predict Molecular Energy using Transformer". ML for Physical Science Workshop on Neurips, 2021.
Page updated
Google Sites
Report abuse