Sept. 22, 2021 — The Lawrence Livermore National Laboratory held its first-ever Machine Learning for Industry Forum (ML₄I) on Aug. 10-12. Co-hosted by the Lab’s High Performance Computing Innovation Center (HPCIC) and Data Science Institute (DSI), the virtual event brought together more than 500 participants from the Department of Energy (DOE) complex, commercial companies, professional societies and academia. Industry sponsors included ArcelorMittal, Cerebras Systems, Ford Motor Company, IBM, Intel, SambaNova Systems, NVIDIA, Intersect360 Research and Rhino Health.

The forum’s goals were to encourage and elucidate the adoption of machine learning (ML) methods for practical outcomes in various industries, particularly manufacturing. Discussions, panel sessions and presentations were organized around three high-level topics: industrial applications, tools and techniques and ML’s impact and potential in industry.

“The forum was created based on LLNL’s interest and experience in helping industry develop artificial intelligence [AI] and ML tools for applications such as manufacturing,” said acting HPCIC director Wayne Miller. Along with the HPCIC, which fosters computing-powered collaborations with the private sector, the Lab’s industry-focused efforts include the High Performance Computing for Energy Innovation (HPC4EI) program and its sub-programs for Manufacturing (HPC4Mfg) and Materials (HPC4Mtls), which leverage the DOE’s HPC facilities to improve energy efficiency and streamline materials development and manufacturing processes. Additionally, LLNL’s Innovation and Partnerships Office (IPO) engages with industry to drive economic growth and brokers commercial licenses and cooperative agreements.

Miller explained: “There is a need to develop collaborations between our data scientists who ‘know how to make ML work’ and industry users who have data, but not much experience in developing ML tools.” DSI director Michael Goldman added: “The DSI’s research and outreach efforts complement the HPCIC’s computational resources and expertise. It made sense to partner with each other on this forum.”

Brenda Ng, an LLNL data scientist who co-organized the event, noted: “My day-to-day work is focused on research and deployment projects. I love applied research, so the forum gave me a chance to hear about others’ experiences and solutions. It also was an outreach opportunity to help industry contacts understand what the Lab does.”

Keynote Variety

Keynote speakers from industry, academia and government kicked off each day’s agenda in turn. Devesh Upadhyay of Ford Motor Company described ML and data-driven approaches to various aspects of vehicle design and manufacturing, including surrogate models and physics-informed neural networks. Pamela Isom, director of the DOE’s Artificial Intelligence Technology Office (AITO), emphasized the importance of improving AI/ML trustworthiness and risk management in cybersecurity and provided an overview of the AITO.

Pieter Abbeel of the University of California, Berkeley’s Robot Learning Lab presented strategies for developing pre-trained neural networks in robotics. The Robot Learning Lab investigates how to make existing AI systems more intelligent as well as how AI can advance science and engineering.

A robot’s brain is a neural network trained to complete tasks it learns from images, text, simulations, demonstrations and other data. Abbeel discussed different types of learning in this context, including multi-task reinforcement learning (RL), unsupervised representation learning, few-shot imitation learning and human-in-the loop RL. “I was excited to share some of the latest advances in AI robotics with a wider audience, as well as my vision as to future research needed in the space,” he said.

Capability Showcase

The event covered the role of HPC, with talks on ML computing workflows, RL for simulations and cognitive simulation. Industry use cases presented to the audience included AI for inspecting defects in steel, computer vision and image processing techniques to automate quality control processes, converged HPC and AI workflows for drug discovery, ML to predict cardiac response to a mitral valve device and uncertainty quantification and surrogate modeling in carbon dioxide capture systems.

LLNL speakers described areas where ML and related data sciences have an impact at the Lab, such as predicting material strength and performance and optimizing manufacturing processes. Computational engineer Vic Castillo presented results from some of his HPC4Mfg projects, which use simulations of critical, energy-intensive manufacturing processes to generate data for ML routines. He stated: “The forum was a great platform to showcase the large variety of ML capabilities at the Lab for a larger industrial audience.”

Castillo’s team has developed fast-running predictive models of computationally expensive simulations that partners can run on gaming desktop computers. He explained: “This empowers the production engineer with good, real-time predictions to help avoid wasting energy and producing low-quality products.” Large-scale simulations can be expensive for private industry, Castillo noted, so “we must be careful to obtain the most useful information from the lowest number of simulation runs.”

Panel Wisdom

The forum featured two panel sessions. The first discussed opportunities for recruiting and training ML talent, integrating them into the workforce and providing resources to develop AI/ML tools. Part of this effort entails bridging the gap between ML taught in classrooms and its practical application in the real world — a goal of the Data Science Summer Institute and other student internship programs around the Lab.

Goldman, who moderated the first panel, stated: “We felt it was crucial to have a workforce conversation at ML₄I. National labs and commercial companies can give students and recent graduates practical opportunities to grow and apply their skills. As employers, we gain staff who are passionate about pushing AI/ML and related areas forward.”

The second panel session considered legal, ethical and cost-benefit challenges in dataset sharing and security — for example, facial recognition or open-source image collection. According to Miller: “Both public and private institutions must grapple with these questions, considering that data is the core resource for all AI/ML development.” Ng added, “This panel opened my eyes to data security and its relevance to the Lab in finding a balance between sharing data and ensuring privacy of customers.”

Common Threads

Overall, Goldman said: “The event highlighted common threads among the participating organizations, and we could have spent more than three days talking about ways to work together.” About 50 presenters answered the call for abstracts in April, and the event’s large turnout and audience response make an ML₄I Forum likely next year.

“The pandemic can make it harder to forge industry partnerships,” noted Ng, who has a new appreciation for how online events are managed. “I think the live and interactive nature of our event, rather than asking speakers to provide pre-recorded video, was more attractive to potential speakers and the audience.”

In addition to Goldman, Miller and Ng, the ML₄I organizing committee included Philip Cameron-Smith and A.J. Simon, HPCIC staff and group leaders in LLNL’s Physical and Life Sciences Directorate; IPO business development executive Charity Follett; and administrators Rosie Aguilar, Florann Mahler and Katie Thomas.

---

Source: Lawrence Livermore National Laboratory