Session A18: Machine Learning for Materials Science I

Predicting aggregate morphology of sequence-defined macromolecules with recurrent neural networks

In this talk, I will present the phase separation behavior of different sequences of a coarse-grained model for sequence defined macromolecules. They exhibit a surprisingly rich phase behavior, and not only conventional liquid-liquid phase separation is observed, but also reentrant phase behavior, in which the liquid phase density decreases at lower temperatures. Most sequences form open phases consisting of aggregates, rather than a normal dense liquid. These aggregates had overall lower densities than the conventional liquid phases and complex geometries with large interconnected string-like or membrane-like clusters. Minor alterations in the sequence may lead to large changes in the overall phase behavior, a fact of significant potential relevance for biology and for designing self-assembled structures using block copolymers. I will discuss recent results from unsupervised manifold learning (UMAP) to classify the different aggregate types and what we can learn from machine learning. Using a bidirectional-Gated Recurrent Units-based Neural Network (RNN), we can now predict which sequence will self-assemble into what aggregate structure.   

Active Learning for Discovering Complex Phase Diagrams with Gaussian Processes

We introduce a Bayesian active learning algorithm that efficiently elucidates phase diagrams. Using a novel acquisition function that assesses both the impact and likelihood of the next observation, the algorithm iteratively determines the most informative next experiment to conduct and rapidly discerns the phase diagrams with multiple phases. Comparative studies against existing methods highlight the superior efficiency of our approach. We demonstrate the algorithm's practical application through the successful identification of the skyrmion phase in the Heisenberg model with antisymmetric interaction.   

Data-driven studies of two-dimensional materials and their nonlinear optical properties

Our research uses data-driven methods to investigate nonlinear optical (NLO) properties in van der Waals (vdW) materials. Utilizing density functional theory calculations and machine learning (ML), we aim to expedite the discovery of stable vdW materials while predicting their crucial second-order susceptibility for NLO phenomena.

We begin by harnessing the Computational 2D Materials Database (C2DB), which comprises 268 non-centrosymmetric, non-magnetic semiconductor monolayers and their corresponding second-order susceptibility tensors across various energy ranges. We data mine C2DB's second-harmonic generation (SHG) spectra to create our dataset.

We use a random forest classifier to determine whether a material exhibits NLO properties. Subsequently, a random forest regression model predicts the second-order susceptibility for three energy ranges (0.4-0.9eV, 0.9-1.6eV, and 1.6-2.1eV). The choice of materials descriptors, derived from atomic properties, influences model performance.

Our research uses ML to shed light on the microscopic origins of NLO phenomena in vdW materials. This study paves the way for the rapid discovery of chemically stable vdW materials capable of second-harmonic generation, with applications spanning photonics and optoelectronics.   

Using Data to Enhance Mechanistic Modeling of Microstructure Evolution in Silicon

Mechanistic models of microstructural evolution in silicon-based materials for electronic and energy applications have been used widely in a variety of settings. Examples include point defect transport and aggregation during bulk silicon crystal growth and wafer processing, surface microstructure evolution during deposition, and impurity segregation at line and planar defects. While much of the physics underpinning these phenomena are well established in principle, uncertainties are often present in the form of unspecified thermophysical model parameters or incomplete descriptions of the physics, especially at the atomistic level. In many cases, these uncertainties can be resolved using data, either measured experimentally or computed using supporting simulations. 

In this talk, two examples of data-assisted mechanistic modeling are presented to illustrate how data may be used to ‘patch over’ modeling elements that are not fully specified. In the first example, a continuum model of oxide precipitation in Czochralski-grown Si crystals is assisted using experimental measurements of precipitate density [1]. While the basic features of oxide precipitation in Si are well-understood, a complete atomistic description of the process is still lacking. In this approach, the model is used to bridge the unknown atomistic properties and the highly coarse-grained experimental measurements. In the second example, we consider the deposition of Ge on an amorphous silica substrate. Here, experimental data is used to first refine an atomistic potential. Then, we use analytical models to make connections between the vast amounts of data generated by the atomistic simulations and experimentally meaningful quantities [2,3].

[1] Y. Yang, A. Sattler, and T. Sinno, Data-Assisted Physical Modeling of Oxygen Precipitation in Silicon Wafers, J. Appl. Phys. 125 (2019), 165705.

[2] C. Y. Chuang, S. M. Han, L. A. Zepeda-Ruiz, and T. Sinno, On Coarse Projective Integration for Atomic Deposition in Amorphous Systems, J. Chem. Phys. 143 (2015) 134703.

[3] C. Y. Chuang, Q. Li, D. Leonhardt, S. M. Han, and T. Sinno, Atomistic Analysis of Germanium on Amorphous SiO2 using an Empirical Interatomic Potential, Surf. Sci. 609 (2013) 221.   

Data-Driven Models for Predicting Stability of Electrocatalysts in Aqueous Environments

There is an increasing demand for alternative catalysts to platinum-group metals for the development of renewable energy technologies such as hydrogen fuel cells. A critical limitation to the performance of fuel cells is the electrochemical durability of the electrocatalytic electrodes. In this work, we implemented a computational framework to identify the factors that define the stability of metal electrodes in aqueous environments. To this end, we employed forward and inverse design with the goal of finding the most stable electrocatalysts using gradient boosting regression and Bayesian optimization. The dataset used to train the machine learning model comprises composition-based features as inputs and electrochemical decomposition energy as output. Maximum electronegativity and mode of covalent radius were identified as the most important features. The model can predict stability with an accuracy of 0.10 eV/atom and an explainability of 98%. This study provides a critical assessment of machine-learning models for estimating the electrochemical stability of catalytic alloys.   

ChemChat | Conversational Expert Assistant in Material Science and Data Visualization

In recent decades, remarkable advancements have been made in the field of computational chemistry and machine learning (ML), yielding a plethora of sophisticated tools and artificial intelligence (AI) models. Despite their potential, these resources have yet to be fully harnessed due to their steep learning curves and their tendency to operate in isolation. Furthermore, the need for capabilities in programming and ML constitute access barriers to the targeted community – often experimental scientists. Concurrently, the advent of large-language models (LLMs) like (Chat)GPT has been revolutionizing various domains. Nevertheless, their efficacy in addressing chemistry-related challenges has been limited. Especially, these models lack knowledge about scientific workflows and the employed operations (e.g. in drug discovery), access to information sources providing up-to-date data, and the ability to accurately reference – but tend to hallucinate in their responses – what questions credibility, trust, and applicability. However, this crucial gap between AI and science can be overcome by integrating task-specific agents into the LLM-powered conversational application and allowing the LLM to reason over their appropriate usage based on provided instructions. It can be anticipated that this will result in a significant increase in the utilization of the developed cheminformatic tools and AI models and contribute to the scientific discovery overall.

Here, we present ChemChat, a web application and conversational assistant with a chatbot-driven user interface that is powered by non-GPT/OpenAI LLMs. Through the integration of existing cheminformatics tools and expert-developed AI models such as PubChem, CIRCA, RDKit, GT4SD, RXN, MolFormer and other knowledge sources the application is capable of assisting scientists in tasks like property calculations, tailored design of molecules, retrosynthesis, forward reaction planning, data visualization, and literature research. Central to the talk we will be demonstrating use case-specific capabilities in comparison to related applications and the architecture and workflow behind ChemChat.   

Towards a Multi-Objective Optimization of Subgroups for the Discovery of Materials with Exceptional Performance

Artificial intelligence (AI) has the potential to revolutionize the design of materials by uncovering correlations and complex patterns in data. However, current AI methods attempt to describe the entire, immense materials space with a single model, while different mechanisms govern the materials behaviors across the materials space. The subgroup-discovery (SGD) approach identifies local rules describing exceptional subsets of data with respect to a given target. Thus, SGD can focus on mechanisms leading to exceptional performance. However, the identification of appropriate SG rules requires a careful consideration of the exceptionality-generality tradeoff. Here, we discuss the notion of SG exceptionality and analyse the tradeoff between exceptionality and generality based on a Pareto front of SGD solutions, thus providing a roadmap for advancing the SGD approach in materials science.   

Equivarient Electron Density Predictions Accelerate Density Functional Theory Calculations

The Hohenberg-Kohn theorem formally maps the ground-state electron density of a many-electron system to the ground-state energy, laying the foundation of modern density functional theory (DFT). Practical applications of DFT start with an estimate charge density, such as the superposition of atomic densities (SAD), and iteratively solve the Kohn-Sham equations until self-consistency. We develop an E(3)-equivarient deep learning model to predict the self-consistent, ground-state electron density that outperform other models found in the literature such as DeepDFT for organic molecules and inorganic materials. Using these predicted electron densities as a starting point for DFT, we show that fewer self-consistent iterations are required to converge a DFT calculation, and that more calculations successfully converge compared to initializing with SAD. In addition, we show that non-self-consistent calculations using the predicted electron densities predict electronic and thermal properties of materials at near-DFT accuracy.

 

DISTRIBUTION STATEMENT A. Approved for public release. Distribution is unlimited.   

Generative neural networks for synthetic PBX microstructures with varying levels of damage to evaluate shock sensitivity through meso-scale simulations

Damage in energetic material (EM) microstructures has the potential to significantly impact the performance of critical national security and safety devices. Understanding and modeling microstructural damage and its effects under various loading conditions requires multi-scale computational models. A significant bottleneck for computational modeling efforts is the paucity of microstructural images at various levels of damage. Here we present HEDS (Heterogeneous Energetic Damage Simulator), a versatile tool designed for generating varying levels of damage within microstructure images of a specific plastic bonded explosive (PBX). The HEDS workflow commences with the preprocessing of available PBX micrographs. Then, leveraging machine learning techniques, HEDS not only removes damage from existing PBX images but also systematically reintroduces damage at different levels or volume fractions, enabling the study of material responses at various damage levels. Furthermore, HEDS offers the capability to create synthetic microstructures, addressing the challenge of limited availability of real microstructure images. HEDS comprises three distinct machine learning models: 1) Microstructure Generation (Diffusion model): enables the generation of a stochastic ensemble of entirely new synthetic microstructures, which is valuable for addressing data scarcity issues in sourcing microstructure images; 2) Image Inpainting (U-Net): this model removes all damage from a given microstructure, creating an undamaged (inpainted) reference; 3) Damage Reintroduction (CycleGAN): reintroduces realistic damage at varying volume fractions into the in-painted microstructure, allowing for controlled assessments of sensitivity to shock loading;. By progressively reintroducing damage into the inpainted microstructure, our framework facilitates a comprehensive analysis of PBX behavior under different levels of damage. HEDS is a useful tool for advancing our understanding of energetic materials and enhancing the safety and effectiveness of related applications.   

Recent Advancements in SISSO as Applied to Thermal Conductivity

Symbolic regression is a promising class of methods for both explainable artificial-intelligence (AI) and materials discovery [1]. The sure-independence screening and sparsifying operator (SISSO) approach [2,3] represents a deterministic way of finding these models as it combines symbolic regression with compressed sensing. Here we present new concepts for the feature creation step that introduce basic grammatical rules for the generated expressions. The utility of these new conditions are demonstrated via toy problems, and then rigorously tested by creating new models for the thermal conductivity of a material [4].

[1] Z. Li and J. Ji arXiv: 2111.12210 (2021)

[2] R. Ouyang et al. Phys. Rev. Mater. 2, 083802 (2018)

[3] T. A. R. Purcell  et al. J. Chem. Phys. 159, 114110 (2023)

[4] T. A. R. Purcell  et al. npj Comput. Mater. 9, 112 (2023)   

Closed-Loop Control of Non-Newtonian Fluid Flow Using Machine Learning

Controlling the flow of non-Newtonian fluids is crucial for advancements in various applications including 3D printing. However, realizing this control necessitates overcoming substantial challenges due to the intricacies in real-time flow monitoring and manipulation. In this research, we introduce a novel methodology that amalgamates electrical flow monitoring with machine learning to meticulously control non-Newtonian fluid flow. We employ a miniaturized non-contact electrical flow monitoring technique based on flow triboelectricity. To realize high-speed, accurate, and real-time control of non-Newtonian flow, we implement a Radial Basis Function Neural Network (RBFNN). The RBFNN is meticulously trained to understand the absolute error and to adapt to the aspired reference rate of the non-Newtonian flow, operating in real-time at a frequency of 1 s^-1. Our machine-learning–enabled closed-loop control system demonstrates exceptional overlap fidelity with the predetermined reference flow rate and performs real-time adjustments to maintain the flow rate. This approach provides immense potential to augment the precision and dependability of various processes that involve non-Newtonian fluids, paving the way for enhanced reliability and accuracy in applications.