Funded Projects

Funded Project Proposals

September 2023

SoCal Climate Change and Environmental Sustainability Program (by CLEWS)

PIs: James Randerson (PI, UCI), Efi Foufoula-Georgiou (co-PI, UCI), Brett Sanders (co-PI, UCI), Jasper Vrugt (co-PI, UCI), Tirtha Banerjee (co-PI, UCI), Michael Goulden (co-PI,UCI), Kristen Davis (co-PI, UCI), Steven Davis (co-PI, UCI), Katherine Mackey (co-PI, UCI).
Partnering National Labs: Los Alamos National Laboratory, Lawrence Berkeley & Los Alamos National Laboratory

SoCal Climate Change and Environmental Sustainability Program (by CLEWS)

PIs: James Randerson (PI, UCI), Efi Foufoula-Georgiou (co-PI, UCI), Brett Sanders (co-PI, UCI), Jasper Vrugt (co-PI, UCI), Tirtha Banerjee (co-PI, UCI), Michael Goulden (co-PI,UCI), Kristen Davis (co-PI, UCI), Steven Davis (co-PI, UCI), Katherine Mackey (co-PI, UCI).

The Climate, Energy, and Water Solutions (CLEWS) Group of UCI is a set of faculty, staff, and students across the Schools of Engineering and Physical Sciences that focus on interdisciplinary research needed to sustainably manage the Earth system. For the 2023-2024 academic year, we propose to build a program at UCI that fosters collaboration between the DOE National Laboratories and SoCal faculty and students in climate change and environmental sustainability. We offer the following program as a service to the UCI campus and the SoCal Hub to strengthen research and educational exchange among the different participating institutions.

Our proposed SoCal Climate Change and Environmental Sustainability Program will include education, workshop, and exchange components in line with the aspirations and goals of the SoCal Hub. These include (1) a seminar series on climate change and environmental sustainability science for the fall quarter, which will engage researchers across SoCal campuses and NLs, (2) a workshop on fire prediction for the winter quarter which will bring together faculty, students and early career researchers across SoCal campuses and NLs, (3) A UC-NL visitor exchange program, and (5). a short course on numerical methods and data analytics in climate science and hydrology for the summer, which will include lecturers and students from across SoCal campuses and NLs.

CLEWS Website

Read the Project Summary

May 2023

Training Robots from Human Videos

PIs: Amit Roy-Chowdhury (Lead PI, UCR), Konstantinos Karydis (UCR), Nikolay Atanasov (UCSD)
Partnering National Lab: Los Alamos National Laboratory

Training Robots from Human Video

PIs: Amit Roy-Chowdhury (Lead PI, UCR), Konstantinos Karydis (UCR), Nikolay Atanasov (UCSD)

Training autonomous systems to work in unstructured environments is a challenging problem and a variety of tools have been proposed for this purpose. Reinforcement Learning (RL), and more recently, Imitation
Learning (IL) are two such methods. Reinforcement learning has shown remarkable success in a variety of fields, including games, robotics, and autonomous driving. However, its effectiveness relies heavily on well designed reward functions, which can be challenging to design. In contrast, imitation learning (IL) offers a data-driven approach to learning behaviors and identifying informative rewards from expert demonstrations,
without requiring explicit reward information.

Visual imitation learning (VIL) has emerged as a promising area of research, specifically focused on learning from high-dimensional visual demonstrations like image sequences or videos. Unlike previous IL works
that tackled low-dimensional inputs, such as proprioceptive features like positions, velocities, and accelerations, VIL deals with more general problems in the physical world. For example, learning to cook by watching videos is a challenge that comes naturally to humans, but remains a difficult task for AI agents.

Existing learning algorithms typically require excessive amounts of curated training data, which cannot be provided by human experts in a realistic work environment, and provide no guarantees on their effectiveness and adaptability across robotic platforms and environments. This project specifically addresses these limitations, by providing algorithms to extract decision and control policies from limited and cross-domain video demonstrations and generalize them across platforms, environments, and tasks

Read the Project Summary

Ensemble-based Robust Models for Computer Vision

PIs: M. Salman Asif (Lead PI, UCR), Amit Roy-Chowdhury (co-PI, UCR) Alfred Chen (co-PI, UCI)
Partnering National Lab: Los Alamos National Laboratory

Ensemble-based Robust Models for Computer Vision

PIs: M. Salman Asif (Lead PI, UCR), Amit Roy Chowdhury (co-PI, UCR) Alfred Chen (co-PI, UCI)

Advances in computer vision and machine learning are largely driven by deep neural networks (DNNs). While these networks can achieve remarkable accuracy on several tasks (e.g., classification, object detection, tracking), they are also vulnerable to distribution shifts, spurious data, and adversarial attacks. In recent years, a number of methods have been proposed to exploit and explain these deficiencies in the DNNs. In a series of recent papers, we developed a number of methods for holistic adversarial attacks on a variety of whitebox and blackbox models that were able to show end-to-end security/safety consequences in both digital- and physical-world settings. In particular, we used vision-language models, context information, and ensemble of surrogate models to generate successful adversarial attacks for blackbox models.

Robustness and safety of autonomous systems depend on the robustness of underlying machine learning and computer vision models. As these technologies become more prevalent and sophisticated, we need to carefully understand the causes of vulnerabilities and ensure their robustness and reliability. Many existing defense-based method rely on adversarial training of a single model against predefined perturbation types, which compromises their performance and offer a false sense of security. In our prior work, we observed that context in the test data and ensemble of models can play an important role in exploiting the vulnerabilities of the victim models. At the same time, the context and ensemble (if used properly) can also provide robustness against attacks and spurious data. In this project, we propose to utilize our expertise to develop robust defense approaches for computer vision and machine learning models.

Read the Project Summary
Seedlings

Seedlings are funded project proposals that expand collaborations initiated at SoCal Hub sponsored workshops.

Data-Driven Many-Body Molecular Dynamics for Predictive Computer Simulations of Complex Molecular Systems

  • Francesco Paesani, UCSD
  • Stephan Mandt, UCI
  • Kipton Barros, LANL

Abstract

Novel machine learning methods for computational chemistry

  • Eric Mjolsness, UCI
  • Kipton Barros, LANL
  • Brenden Petersen, LLNL

Abstract

A machine learning enhanced multiscale modeling framework for plasticity and failure in heterogeneous materials

  • Shabnam Semnani, UCSD
  • Jaime Marian, UCLA
  • Diane Oyen, LANL
  • Joshua White, LLNL
  • Eric Herbold, LLNL
  • Michael Homel, LLNL

Abstract

Design of nanostructured bistable materials as a case study for complex systems center proposals

  • Nicholas Boechler, UCSD
  • Alicia Kim, UCSD
  • Lorenzo Valdevit, UCI
  • Christopher Spadaccini, LLNL

Abstract

Navigating the combinatorial alloy chemistry challenge of advanced multicomponent materials for nuclear energy applications

  • Jaime Marian, UCLA
  • Amartya Banerjee, UCLA
  • Mathieu Bauchy, UCLA
  • Morris Wang, UCLA
  • Sergei Krasheninnikov, UCSD
  • Stuart Maloy, LANL
  • Osman El Atwani, LANL
  • Saryu Fensin, LANL

Abstract

Southern California hub for the electron-Ion collider (EIC) physics: gearing up for the EIC Era

  • Zhongbo Kang, UCLA
  • Huan Huang UCLA
  • Miguel Arratia, UCR
  • Cesar da Silva, LANL
  • Astrid Morreale, LANL
  • Ron Stoltz, LLNL
  • Aaron Angerami, LLNL
  • Ramona Vogt, LLNL

Abstract

Two-component unconventional superconductors and topological states of matter

  • Stuart Brown, UCLA
  • Mark Sherwin UCSB
  • Filip Ronning, LANL

Abstract

Project HUNTER (Heavy Unseen Neutrinos from Total Energy-momentum Reconstruction)

  • Alexander Kusenko, UCLA
  • Graciela Gelmini, UCLA
  • Alexander Kusenko, UCLA
  • Paul Hamilton, UCLA
  • Eric Hudson, UCLA
  • Kevork Abazajian, UCI
  • George Fuller, UCSD
  • Daniel Alves, LANL
  • Christopher Fryer, LANL

Abstract

Data-Driven Many-Body Molecular Dynamics for Predictive Computer Simulations of Complex Molecular Systems

Two of the most challenging problems at the intersection of electronic structure theory and molecular dynamics simulations are the accurate representation of intermolecular interactions and the development of efficient algorithms applicable to large systems. To some extent, these two problems are antithetical, since accurate calculations of molecular interactions typically require correlated electronic structure methods that are computationally too expensive for applications to large systems. To overcome existing barriers that preclude predictive, molecular-level computer simulations of generic molecular systems, we are developing a novel theoretical and computational methodology, denoted as data-driven many-body molecular dynamics (MB-MD), that integrates data-driven models, which are machine-learned to represent quantum mechanical interactions, with physics-based models, which represent many-body interactions. Through the implementation of MB-MD on exascale high-performance computing, platforms, we aim to enable predictive, large-scale computer simulations of complex molecular systems, with specific on those relevant for environmental, energy, and biomedical applications.

Novel machine learning methods for computational chemistry

We seek to combine numerical machine learning with symbolic algebraic methods in pursuit of physics-based computational chemistry. This ambition suggests problem representations including graphs for molecular structure, deep neural networks (also graph structures) for function approximation under continuous optimization, and syntax trees for domain-appropriate symbolic mathematical expressions using reinforcement learning.

Machine-learnable computational chemistry models have been formulated in terms of combinations of antisymmetric determinants and/or many-body energy functions, incorporating neural networks and constrained by domain-specific graph structure. The resulting graph-structured models can be searched over by reinforcement learning on symbolic expressions that include composition operators defined over “graph lineages”, regular families of graphs of increasing size that may approach a continuum limit object [Scott and Mjolsness, 2021]. These composition-operators will be key elements in a new formal language of models, designed to obey useful algebraic identities. The resulting nonstandard approach provides the possibility of unexpected breakthroughs.

[Scott and Mjolsness, 2021] “Graph diffusion distance: Properties and efficient computation”, C.B. Scott and Eric Mjolsness. PLOS One, https://doi.org/10.1371/journal.pone.0249624  April 27, 2021.

A Machine Learning Enhanced Multiscale Modeling Framework for Plasticity and Failure in Heterogeneous Materials

Despite progress made in discovering mechanisms behind dissipative deformations and failure in materials, many questions remain unanswered regarding the sequence and interaction of various deformation and failure mechanisms. In particular, accurate modeling and prediction of initiation and evolution of damage and fracture remains a major challenge. This is mainly because processes contributing to material failure span across multiple scales, starting from the atomistic all the way to the macroscopic scale. This pathway project aims to leverage Machine Learning techniques to effectively link atomic/nano-scale processes to the larger scales for accurate prediction of nonlinear and path-dependent deformation in materials. The proof of concept will be shown with simple 1D/2D geometries of materials such as metallic alloys and polycrystalline materials. This framework will be informed and validated using material imaging and experiments.

Design of nanostructured bistable materials as a case study for complex systems center proposals

Complex systems are ubiquitous in our natural and engineered world, and typically exhibit “emergent phenomena” wherein nonlinear interplay between system components yields behavior qualitatively different to the constituent parts. Complex nonlinear systems are often seen as a liability because they are vastly more difficult to predict than linear systems, and the emergent phenomena they support can be detrimental. However, the downside of emergent phenomena should not be allowed to block access to their astounding beneficial possibilities, wherein the performance of systems can be increased dramatically. This project explores a case study in the form of impact mitigation via highly nonlinear nanostructured bistable materials, to serve as a key example in center proposals. This includes exploring the performance of single-material-construction and nanoscale manufacturing of bistable lattice materials, and design of bistable structural elements that suppress solitary wave transmission.

Navigating the combinatorial alloy chemistry challenge of advanced multicomponent materials for nuclear energy applications

In a carbon-free economy, nuclear power will surely play a fundamental role as a clean and cost-competitive energy source. However, new-generation nuclear concepts will involve temperature and irradiation conditions for which no suitable materials presently exist. A promising class of materials that has shown great potential in addressing the issues associated with these conditions are the so-called high entropy alloys (HEA), a new type of concentrated multicomponent metallic alloys that combines four or more transition metal elements in equiatomic proportions. Coming up with suitable alloy compositions is thus a metals chemistry problem of combinatorial proportions. Among the tools that can be called upon to solve this advanced materials development grand challenge are high-throughput computer simulation –including preparations for the quantum-computing era–, advanced manufacturing techniques based on scalable 3D printing, and ultraprecision characterization techniques. Here we use these approaches in conjunction, identifying and taking advantage of potential synergisms to narrow down the parameter space and expedite the selection of viable materials. .

Southern California hub for the electron-Ion collider (EIC) physics: gearing up for the EIC Era

The science of the future Electron-Ion Collider (EIC) will be at the heart of the U.S. nuclear physics program in the next several decades. A major pillar of the EIC scientific program is the discovery of a new regime of nuclear matter, known as color glass condensate, dominated by a highly dense and saturated system of gluons. The quest to achieve such a mission requires the collective and coordinated effort of experts in nuclear and particle physics. We have established a UC southern hub for the EIC physics - through a synergy between theory and experiment, we will assess the potential for the implications of Quantum Chromodynamics to the science of the EIC. Our hub will develop new theoretical and phenomenological tools, will perform experimental and detector simulations that help further assess the feasibility of these measurements, and will design, prototyping and construction of EIC calorimeter detectors.

Two-component unconventional superconductors and topological states of matter

The project consists of two parts. The first is a collaborative effort between the research groups at UCLA (Stuart Brown) and at LANL (Filip Ronning, Priscila Rosa, Michihiro Hirata). The research team is focused on a compelling problem in the field of unconventional superconductivity, namely the stabilization and nature of states comprised of two component order. Its occurrence is rare, and until recently thought to occur only under the symmetry-imposed degeneracy condition. It also serves as the paradigmatic platform for strong topological superconductivity. The second component is the organization of workshops in quantum materials organized by the 3 Principal Investigators named here: Stuart Brown, Mark Sherwin (UCSB), Filip Ronning. The objective is to launch a nucleus for scientific exchange that is both intercampus, and between the campuses and Los Alamos National laboratory. The workshops are hosted by the Institute for Materials Science at LANL, which Dr. Ronning heads, and the first of these, to be held August 1-5, 2022, is focused on Topology and correlations in frustrated lattices.

Project HUNTER (Heavy Unseen Neutrinos from Total Energy-momentum Reconstruction)

The HUNTER Southern Hub will help launch an unprecedented effort to search for sterile neutrinos, which can account for all or part of dark matter and whose existence is suggested by the neutrino masses.  The Hub will bring together theoretical physicists from UCLA, LANL, UCI, and UCSD to lay the groundwork for the HUNTER Experiment (Heavy Unseen Neutrinos from Total Energy-momentum Reconstruction), which has already received seed funding ($1.2M)  from the Keck Foundation and is seeking full funding from the NSF. Excitingly, the apparatus developed under the HUNTER program will not only allow the tightest laboratory constraints (or discovery!) of sterile neutrinos, but can also be used for a host of other fundamental physics measurements.  As such, this seed grant will be used to establish the HUNTER facility, which will be used by the Southern California UC team in partnership with LANL for at least the next two decades.