AI for Big Data Problems in Advanced Imaging, Materials Modeling and Automated Synthesis: AI and ML for Materials Discovery I
Sponsored by: TMS Advanced Characterization, Testing, and Simulation Committee
Program Organizers: Mathew Cherukara, Argonne National Laboratory; Subramanian Sankaranarayanan, University of Illinois-Chicago; Badri Narayanan, University of Louisville
Monday 8:00 AM
October 10, 2022
Room: 304
Location: David L. Lawrence Convention Center
Session Chair: Badri Narayanan, University of Louisville; Mathew Cherukara, Argonne National Laboratory
8:00 AM
A Feasibility Study of Machine Learning-assisted Alloy Design Using Wrought Aluminum Alloys as An Example: Yasaman Jamalipour Soofi1; Md Asad Rahman1; Yijia Gu1; Jinling Liu1; 1Missouri University of Science and Technology
Machine learning (ML) typically requires large datasets for reliable predictions, which may not be realistic for most commercial alloy systems. Also, the alloy development requires a full set of balanced properties, most of which have not been studied by ML yet. In this study, we focused on the practicality and reliability of ML in alloy design using commercial wrought aluminum alloys as an example. The dataset used in this study contains 236 entries and 15 alloy properties. We systematically evaluated various ML models with a focus on the bias-variance trade-off. We further explored the possibility of engineering the feature space to improve the ML models. Lastly, our feature importance analysis suggested the soundness of the developed models and provided new insights into the underlying processing-structure-property relations. This study demonstrated that it is feasible to use machine learning and data mining techniques to assist the alloy design using realistic small datasets.
8:20 AM
Are Process-Structure-Property Relationships Useful for Materials Design?: Raymundo Arroyave1; 1Texas A&M University
The focus of goal-oriented materials design is to find the necessary chemistry/processing conditions to achieve the desired properties. In this setting, a material’s microstructure is either only used to carry out multiscale simulations to establish an invertible quantitative process-structure-property (PSP) relationship, or to rationalize a posteriori the underlying microstructural features responsible for the properties achieved. The materials design process itself, however, tends to be microstructure-agnostic. In this talk, we attempt to resolve the issue of whether ‘PSP’ is a superior paradigm for materials design in cases where the microstructure itself cannot be (directly) manipulated to optimize materials’ properties. To this end, we formulate a novel microstructure-aware closed-loop multi-fidelity Bayesian optimization framework for materials design and rigorously demonstrate the importance of the microstructure information in the materials design process.
8:40 AM
A.I. Driven Sustainable Aluminum Alloy Design: Fatih Sen1; 1Novelis
Aluminum has been increasingly the sustainable material of choice for aerospace, automotive and beverage cans due to its high strength-to-weight ratio and infinite recyclability. Aluminum recycling is expected to surge in the coming years, and more end of automotive life aluminum scrap will be available in the market. It is of great importance to design aluminum alloys that can consume such scrap to minimize prime aluminum use, and simultaneously satisfy the performance requirements for desired applications, such as strength, formability, corrosion, etc. In the present work, we have developed a physics-informed machine-learning framework coupled to a scrap flow and blending model to accelerate sustainable aluminum alloy design workflows. We have used computational thermodynamics methods to estimate microstructural features pertaining to strength, formability and corrosion properties and integrated them into our materials informatics framework. Our model predictions for high recycle content alloy chemistries was then validated through lab trials.
9:00 AM
De Novo Molecular Drug Design Using Deep and Reinforcement Learning: Srilok Srinivasan1; Rohit Batra2; Henry Chan2; Mathew Cherukara2; Jonathan Steckbeck1; Nicholas Nystrom1; Subramanian Sankaranarayanan2; 1Peptilogics; 2Argonne National Laboratory
The astronomically large chemical space of drug discovery exceeds what can be efficiently explored by current screening approaches such as docking or high-throughput library screening, which are limited by high computational cost or the low coverage of libraries over the chemical space. We present a de novo design strategy that leverages deep learning and reinforcement learning to design compounds that effectively bind to a target protein of interest. Our workflow integrates a Monte Carlo Tree Search algorithm with deep neural network architectures to operate in a generative fashion and effectively sample the vast chemical space. We generate several new biomolecules that outperform or show competitive performance compared to, known FDA-approved and non-FDA-approved biomolecules from existing databases.
9:20 AM
High-dimensional Neural Network Potential for Liquid Electrolyte Simulations: Applications to Li-ion Battery Materials: Garvit Agarwal1; Steven Dajnowicz1; James Stevenson1; Leif Jacobson1; Farhad Ramezanghorbani1; Karl Leswing1; Mathew Halls1; Robert Abel1; 1Schrodinger Inc
To move towards accurate and reliable modeling of Li-ion battery (LIB) chemistries, we developed a machine-learned potential for liquid electrolyte simulations. The potential was constructed using the charge recursive neural network architecture, which includes both long-range interactions and global charge redistribution. The potential uses non-periodic (cluster) DFT training data, allowing the use of more accurate functionals, like the range-separated hybrid ωB97X-D3BJ, which would be prohibitively expensive for generating datasets with periodic DFT. Here, we focus on seven carbonate solvents and LiPF6 salt in the LIB technology. Despite only training to cluster data, the predicted bulk thermodynamic properties and transport properties are in excellent agreement with experiments. The potential reproduces the concentration and temperature dependence for viscosity and diffusivity of ions and solvent. Furthermore, we demonstrate the capability of the model to accurately predict the solvation structure of ions using a comparison of the radial distribution functions with experimental data.
9:40 AM Question and Answer Period