Day 1: Methods and Algorithms

Day 1

April 20th, 2021

Methods, Math, and Algorithms

All times are in PT

08:00 - 10.00

Introductory, Big-Picture, and Integrated-Perspective Talks

Chairs: Marcus Noack, Petrus Zwart

08:00 - 08:15

Dr. Marcus M. Noack: Introduction to the Workshop and to Autonomous Discovery

08:15 - 08:30

Prof. James A. Sethian: The Center for Advanced Mathematics for Energy Research Applications (CAMERA)

UC Berkeley

08:30 - 09:00

Dr. Draguna Vrabie

Pacific Northwest National Laboratory

Autonomous control systems and learning

Autonomous control systems are those that accomplish their objectives in real-world scenarios and under uncertainty in the environments and can compensate for changes and failures without external intervention. In this talk, I will provide a brief overview of modern controls systems and their transition to incorporate data-driven knowledge. I will then introduce a data-driven predictive control approach that uses physics-informed deep learning representations for synthesizing control policies and models of the unknown dynamic systems.

09:00 - 09:30

Dr. Benji Maruyama

Air Force Research Laboratory


Autonomous Research Systems for Materials Development


The current materials research process is slow and expensive; taking decades from invention to commercialization. Researchers are now exploiting advances in artificial intelligence (AI), autonomy & robotics, along with modeling and simulation to create research robots capable of doing iterative experimentation orders of magnitude faster than today. We will discuss concepts and advances in autonomous experimentation in general, and

associated hardware, software and autonomous methods.

We propose a “Moore’s Law for the Speed of Research,” where the rate of advancement increases exponentially, and the cost of research drops exponentially. We consider a renaissance in “Citizen Science” where access to online research robots makes science widely available. This presentation will highlight advances in autonomous research and consider the implications of AI-driven experimentation on the materials landscape.


09:30 - 10:00

Dr. Kevin G. Yager

Center for Functional Nanomaterials, Brookhaven National Laboratory

Autonomous Analytics and Control in X-ray Scattering

Autonomous control of complex scientific instruments--such as synchrotron beamlines--holds enormous promise for accelerating experiments and materials discovery. This talk will discuss the vision for autonomous experiments (AE) using a synchrotron x-ray scattering beamline as the key example. Deep learning is used to classify x-ray detector images, with performance improving when domain-specific data transformations are applied. To close the autonomous loop, we deploy a general-purpose algorithm based on gaussian processes. The underlying modeling can exploit system-specific information, while the objective function can be tailored to a particular experiment, balancing knowledge gain, cost, and search targets. Examples from recent autonomous experiments will be presented, including measuring nanoparticle ordering, 3D-printed materials, multi-dimensional combinatorial libraries, and real-time photo-thermal processing.

10:00 - 10:30

Break

10:30 - 12:30

Diving Deep into Methods, Math, and Algorithms

Chairs: Petrus Zwart, Marcus Noack

10:30 - 11:00

Prof. Kristofer Reyes

Department of Materials Design and Innovation, University of Buffalo

From Black-Box to Problem-Fluency: An overview of models and algorithms for Closed-Loop, Autonomous Materials Design

Closed-loop materials design platforms have grown in recent popularity, with many robot scientists coming online in the past few years. Key to their operation are the robots' "brains" -- the machine learning models and decision-making algorithms they use to explore experiment space without human intervention. Many generic black-box models have enabled the rapid development of prototype platforms that have shown the promise of autonomous materials development. While at a nascent state, the field is primed to explore more sophisticated, problem-fluent methods and models, hoping that such practices can further accelerate discovery at scale. We will walk down the path from such black-box models to more problem-aware ones in this talk, highlighting core concepts and seminal works and motivating the practical need for more complex models. We place many of the topics covered in this workshop's discussions and tutorials in context with one another and explore promising new directions to modeling materials experiments and broader systems, all in the context of autonomous discovery.

11:00 - 11:30

Dr. Marcus M. Noack

Computational Research Division, Lawrence Berkeley National Laboratory

Domain-Aware Gaussian Processes and High-Performance Mathematical Optimization for Optimal and Autonomous Data Acquisition

Gaussian Processes have shown to be a powerful tool for autonomous control of data acquisition due to their robustness, analytical tractability, and natural inclusion of uncertainty quantification. In this talk, I want to present our work on a general, flexible, and powerful GP-driven framework for autonomous data acquisition. The focus will lie on making a Gaussian process domain aware, how this awareness can be used for decision-making, and the challenges that come with domain awareness.

11:30 - 12:00

Dr. Francis J. Alexander

Computational Science Initiative, Brookhaven National Laboratory

Optimal Experimental Design via Mean Objective Cost of Uncertainty

I will discuss the concept of the Mean Objective Cost of Uncertainty (MOCU) in the context of optimal experimental design. The method will be described in detail, including variants and new directions, I will then describe how this concept is currently being used in the resource allocation and experimental design first for simple dynamical systems, then complex biological systems, and finally, in the context of ExaScale computing.

12:00 - 12:30

Dr. Maria K. Chan

Center for Nanoscale Materials, Argonne National Laboratory

Integrating Theory and Modeling in Autonomous Discovery

Achieving scientific understanding from high throughput and autonomous experimental frameworks often requires the incorporation of modeling and simulation data. We will discuss the relevant challenges for integrating theory and experiment using AI, and the roles played by atomistic and first principles modeling in deciphering characterization data for autonomous feedback and continuous space search for materials design.

12:30 - 14:00

Lunch Break

Open Zoom Room for Lightning Talks and Networking

14:00 - 16:30

Breakout Sessions: Methods, Math, Algorithms, and Infrastructure

Basics and Applications of Gaussian Processes

Chairs: Hector Garcia Martin, Gilad Kusne

In Concert:

Austin McDannald, NIST

Gilad Kusne, NIST

Intro to GPs

Hands-On with GP Tools

Hands-On with Active Learning


Basics and Applications of Reinforcement Learning

Chairs: Kevin Yager, Petrus Zwart

14:00 - 14:30: Rama Vasudevan, ORNL

Reinforcement learning in materials science: problems, considerations, and applications to synthesis, materials design, and microscopy.

14:30 - 15:00: Auralee L. Edelen, SLAC

Neural Network Based Control Applications in Particle Accelerators

15:00 - 15:30: Jean Betterton, SU

Reinforcement Learning for
Adaptive Illumination with X-rays

15:30 - 16:00: Pankaj Rajak, ANL

Autonomous Material Synthesis with Offline Reinforcement Learning




Basics and Applications of Neural Networks

Chairs: Dani Ushizima, Alex Hexemer

14:00 - 14:30: Maxim Ziatdinov,ORNL

Deep learning for imaging and spectroscopy: Model training, evaluation, and prediction under dataset shift.

14:30 - 15:00: Daniela Ushizima,LBNL

Neural Networks for Scientific Images

15:00 - 15:30: Nathan Melton, LBNL

Neural Networks - advancing scientific discovery and data analysis

15:30 - 16:00: Daan Pelt, Leiden U.

Optimizing neural networks for scientific image applications




Data Analytics & Infrastructure

Chairs: Nicholas Schwarz, Martin Boehm

14:00 - 14:30: June Lau, NIST

Construction of an electron microscopy data ecosystem: the NexusLIMS project

14:30 - 15:00: Ryan L. Chard, ANL

Gladier: The Globus architecture for data-intensive experimental research

15:00 - 15:30: Thomas Caswell, BNL

Bluesky Adaptive: Bringing AI to production

15:30 - 16:00: Zachary Trautt, NIST

Navigating the Materials Data Landscape






Uncertainty Quantification

Chair: Apurva Mehta, Logan Ward

14:00 - 14:30: Prof. Claudia Schillings, U. Mannheim

Introduction to UQ


14:30 - 15:00: Mike McKerns, UQ F.

mystic: software for autonomous discovery, design, and control under uncertainty

15:00 - 15:30: Aashwin Mishra, SLAC

Uncertainty Quantification and Out Of Distribution Robustness For Deep Neural Networks

15:30 - 16:00: Shengjia Zhao, SU

Transforming any predictions into calibrated probabilities with conformal calibration

Mathematical Optimization

Chairs: Bobby Sumpter, Sergei Kalinin

14:00 - 14:30: Willie Neiswanger, SU

Going Beyond Global Optima with Bayesian Algorithm Execution


14:30 - 15:00: David Perryman, LBNL

A Hybrid-Local-Deflated-Global Optimizer

15:00 - 15:30: Juliane Mueller, LBNL

Surrogate Optimization of HPC Applications

15:30 - 16:00: Jeff Donatelli, LBNL

Exploiting Mathematical Structure in Inverse Problems via Multi-Tiered Iterative Projections





Emerging Mathematics for Autonomy

Chairs: Reeja Jayan, Simon Billinge

14:00 - 14:30: Roger Ghanem, USC

Prediction and adaptation as constrained statistical inference

14:30 - 15:00: José Miguel Hernández-Lobato, Cambridge

Successor Uncertainties: Exploration and Uncertainty in Temporal Difference Learning

15:00 - 15:30: Shreyas Sundaram, Purdue U.

Enabling Inference by Distributed Autonomous Agents

15:30 - 16:00: Evangelos Theodorou, Georgia Tech


Decision Making Under Uncertainty: From Robotics and Autonomous Systems to Applied Physics


16:30 - 17:00

Wrap-Up