Schedule & Abstracts

Laura Anderson, National Science Foundation
Funding Opportunities for AI-Assisted Chemical Synthesis at the NSF

Recent and current funding opportunities for AI-assisted synthesis in the Division of Chemistry will be discussed with highlights from some of these awards.

Maria Chan, Argonne National Lab
Integrating Modeling and Experimental Characterization Data for Accelerated Materials Understanding

The use of high throughput computation and computational database has led to significant acceleration in materials design and discovery. For in-depth understanding of functional materials structure and operando changes thereof, however, multimodal characterization utilizing x-ray, electron, scanning probe, etc. is paramount. This talk will discuss how computational and experimental data can be simultaneously utilized, with the help of AI/ML approaches, to accelerate understanding. Issues surrounding FAIR data, especially for experimental characterization, will also be discussed.

Dogus Cubuk, Google DeepMind
Scaling Deep Learning for Materials Discovery

Despite the recent advances in physical simulations and machine learning, the exploration of novel inorganic crystals remains constrained by the expensive trial-and-error approaches. Recent developments in deep learning have shown that models can showcase emergent predictive capabilities with increasing data and computation in fields such as language, vision, and biology. In this talk, I will present our recent results on how graph networks trained at scale can reach unprecedented levels of generalization, improving the efficiency of materials discovery by an order of magnitude. Building on the 48,000 stable crystals identified in ongoing studies, improved efficiency enables the discovery of 2.2 million stable structures with respect to the current convex hull, many of which had escaped prior human chemical intuition. The scale and diversity unlock surprising modeling capabilities for downstream applications, leading in particular to highly accurate and robust learned interatomic potentials that can be used in condensed-phase molecular dynamics simulations and high-fidelity zero-shot predictions.

Elif Ertekin, UIUC
Quantifying Uncertainty when Applying Data Science and Probabilistic Modeling to Materials: Towards Actually Knowing What We’re Doing

Realizing the promise of machine learning applied to materials requires models that are robust, transferrable, and that can be evaluated in rigorous ways.  It requires knowing when and where the model is useful, and when and where the model may perform poorly. Machine learning surrogates face several distinct challenges when applied to the materials realm including noisy data, small data, label imbalances, and others. While uncertainty quantification itself is an established field with rigorous methods, in this presentation I will share my perspective on the unique challenges that arise when quantifying the uncertainty of statistical models for materials domain applications.  We will look at examples including (i) accounting for the uncertainties in ground truth data itself for materials discovery, (ii) evaluating model uncertainties in surrogate models for property prediction, and (iii) assessing uncertainties of machine learned interatomic potentials in the presence of hard-to-detect covariate distribution shifts. 

Rafael Gomez-Bombarelli, MIT
Computational Materials Design with Machine Learning and Atomistic Simulations

Designing new materials is vital to address pressing challenges in health, energy, and sustainability. The computational techniques of atomistic simulation and machine learning (ML) offer an avenue to rapidly invent new materials and navigate this enormous space. By populating the continuum between physics-based simulations and machine learning, the Learning Matter Lab seeks to enable rapid, computation-first design of materials that accelerate the materials discovery cycle. 

ML enables a new paradigm: surrogate models fuse simulations with experimental data at a fraction of the cost, while embedding physics-based priors to ensure robustness and transferability. The speed and accuracy of these models allows inverse design of materials for many applications. We will present our progress in enabling experimentally validated materials design for multiple materials classes and applications, such as heterogeneous nanoporous thermo-catalysts; composition and surface engineering of oxide electrocatalysts, organic and inorganic solid electrolytes for batteries, therapeutic peptides, recyclable thermoset plastics, carbon-capture nucleophiles to replace amines, stable and optically tuned organic electronics for OLED and OPD. 

Vincenzo Lordi and Fei Zhou, Lawrence Livermore National Laboratory
Identifying and Quantifying Uncertainty in Fitted Interatomic Potentials for Molecular Dynamic

A grand challenge in materials simulations is to predict macroscopic properties within estimated uncertainty bounds without relying on any calibration data. Recently, machine learned potentials that balance computational efficiency and accuracy by fitting against a ground truth theory have enabled the prediction of macroscopic materials properties from extreme-scale molecular dynamics simulations. However, even state-of-the-art methods fall short in estimating the effects of model selection, fitting approaches, and train/test set selection toward simulation outcomes and their associated uncertainties. Here, we discuss needs and approaches for improving the rigor of uncertainty estimates relative to a ground truth theory, thereby enabling property predictions with quantified uncertainty. We highlight the gaps in defining disparate sources of model uncertainty against the desired ground truth theory, which are required to fully automate the process of developing, testing, and deploying reliable potential models. In particular, we focus on large-scale simulations where experimental validation is challenging and where theoretical validation against ground truth is intractable. Brief comments on specific problems—and opportunities—that arise when applying these approaches to extreme-scale simulations, e.g., billions of atoms, also will be presented. 

Prepared by LLNL under Contract DE-AC52-07NA27344, and funded by the LDRD program at LLNL under project tracking codes 23-SI-006 and 22-ERD-055.

John Schlueter, Program Director DMREF, NSF
The Designing Materials to Revolutionize and Engineer Our Future (DMREF) Program at the NSF

The Materials Genome Initiative (MGI) is a multi-agency partnership that seeks to accelerate the progression of materials research across the Materials Development Continuum for the benefit of society. By coupling a predictive computationally led and data-driven approach with experimental synthesis and validation via an iterative feedback loop, MGI promotes the rapid design, discovery, development, and deployment of advanced materials that will ensure sustained American leadership in sectors including clean energy, national security, and human welfare. After a decade of progress that has witnessed a paradigm shift in philosophy of materials research, the second MGI Strategic Plan was released in 2021 defining three primary goals for the next five years:  1) Unifying the Materials Innovation Infrastructure, 2) Harnessing the Power of Materials Data, and 3) Educating, Training, and Connecting the Materials Research and Development Workforce. The Designing Materials to Revolutionize and Engineer our Future (DMREF) program at the National Science Foundation (NSF) partners with other federal agencies to promote these objectives. DMREF includes participation from ten divisions in four directorates at the NSF to address fundamental materials discovery and development. Additional information about DMREF can be found at DMREF.org. This talk will provide an overview of the DMREF program and data-related funding opportunities sponsored by NSF.   

Jennifer Schumacher, 3M
Grand Challenges for Data-Driven Materials Science in Industry

Data-driven materials science in industry offers several grand challenges for balancing customer needs with scale, cost, and raw materials. In the case of materials design or discovery, practical results need to also consider supply chain availability, potential raw material substitutions, and sustainability. Often the next challenge is the standard characterization of a material may not relate to the product level performance metric and often requires multi-modal properties (e.g., optically clear with particular bond strength). Additional complexities are introduced when considering different process possibilities and the interactions with other materials in the final product. While there is a plethora of problems to solve, it is an exciting time to make significant impact on the next generation of materials science. 

Francesca Tavazza, Materials Measurement Science Division, National Institute of Standards and Technology
Prediction Intervals Determination for Regression Models of Atomistic Quantities

Uncertainty quantification in AI-based predictions of material properties is of immense importance for the success and reliability of AI applications in material science. While confidence intervals are commonly reported for machine learning (ML) models, prediction intervals, i.e. the evaluation of the uncertainty on each prediction, are more rarely available. In this work we investigate several UQ approaches to determine such prediction intervals for atomistic quantities of interest for the ICME community. We identify each approach’s advantages and disadvantages and compare their results. All data for training and testing were taken from the publicly available JARVIS-DFT database, and the codes developed for computing the prediction intervals are available through JARVIS-tools.

Fan Zhang, University of Texas at Dallas
Deep Optical Sensing with Emergent Materials

I will discuss how micron-scale moiré graphene can encode all the information of light through nonlinear light-matter interactions, and how convolutional neural network models can decode power, wavelength, polarization state, etc. simultaneously and instantaneously, yielding a groundbreaking technology: intelligent infrared sensing. I will also discuss how quasi-1D topological materials can participate in the next phase.